Jean-Franois Magni, Samir Bennani and

Jan Terlouw (Eds.)

Robust Flight Control:

A Design Challenge

i
This book was rst printed by Springer-Verlag, 1997
Lexture Notes in Control and Information S ien es, 224.

Editors
Jean-Franois Magni, Do teur s S ien es
ONERA CERT, Dpartement d'tudes et Re her hes en Automatique,
BP 4025, F31055 Toulouse Cedex, Fran e.

Samir Bennani, Ir.

Delft University of Te hnology, Fa ulty of Aerospa e Engineering, Kluyver-
weg 1, 2629 HS Delft, The Netherlands.

Jan Terlouw, Ir.

National Aerospa e Laboratory NLR, Flight Me hani s Department, Anthony
Fokkerweg 2, 1059 CM Amsterdam, The Netherlands.

ii
 ROBUST FLIGHT CONTROL:
A DESIGN CHALLENGE

EDITORS

J.-F. Magni, S. Bennani & J. Terlouw

GARTEUR ACTION GROUP FM(AG08)

Resear h Establishments:
Centro Italiano Ri er he Aerospaziali (CIRA, Italy),
Deuts he Fors hungsanstalt fr Luft- und Raumfahrt (DLR, Germany),
Defen e Resear h Agen y (DRA, United Kingdom),
Instituto Na ional de T ni a Aeroespa ial (INTA, Spain),
Laboratoire d'Automatique et d'Analyse des Systmes (LAAS, Fran e),
National Aerospa e Laboratory (NLR, The Netherlands),
O e National d'Etudes et de Re her hes Arospatiales (ONERA, Fran e).

Industry:
Alenia Aeronauti a (ALN, Italy),
Avro International Aerospa e (AVRO, United Kingdom),
British Aerospa e, Dynami s (BAe-D, United Kingdom),
British Aerospa e, Military Air raft (BAe-MA, United Kingdom),
Cambridge Control Ltd (CCL, United Kingdom),
Daimler-Benz Aerospa e Airbus (DASA, Germany),
Fokker Air raft Company (FAC, The Netherlands),
Saab Military Air raft (SMA, Sweden).

Universities:
Craneld University (CUN, United Kingdom),
Delft University of Te hnology (DUT, The Netherlands),
Linkping University (LiTH, Sweden),
Loughborough University (LUT, United Kingdom),
University of Cambridge (UCAM, United Kingdom),
University of Lei ester (ULES, United Kingdom),
Universit di Napoli Frederi o II (UNAP, Italy),
Universitad Na ional de Edu a in a Distan ia (UNED, Spain).

iii
iv
Contents

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

1 Introdu tion. Jan Terlouw and Chris Fielding . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

Tutorial part
2 Multi-Obje tive Parameter Synthesis (MOPS). Georg Grbel and Hans-Dieter
Joos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

3 Eigenstru ture Assignment. Lester Faleiro, Jean-Franois Magni, Jess M.

de la Cruz and Stefano S ala . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

4 Linear Quadrati Optimal Control. Fran es o Amato, Massimiliano Mattei

and Stefano S ala . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

5 Robust Quadrati Stabilization. Germain Gar ia, Ja ques Bernussou, Jamal

Daafouz and Denis Arzelier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

6 H1 Mixed Sensitivity. Mark R. Tu ker and Daniel J. Walker . . . . . . . . . . . 52

7 H1 Loop Shaping. George Papageorgiou, Keith Glover, Alex Smerlas and
Ian Postlethwaite . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

8 -Synthesis. Samir Bennani, Gertjan Looye and Carsten S herer . . . . . . . 81

9 Nonlinear Dynami Inversion. Binh Dang Vu . . . . . . . . . . . . . . . . . . . . . . . . . 102

10 Robust Inverse Dynami s Estimation. Ewan Muir . . . . . . . . . . . . . . . . . . . 112

11 A Model Following Control Approa h. Holger Duda, Gerhard Bouwer,

J.-Mi hael Baus hat and Klaus-Uwe Hahn . . . . . . . . . . . . . . . . . . . . . . . . . 116

12 Predi tive Control. Jan Ma iejowski and Mihai Huzmezan . . . . . . . . . . . 125

13 Fuzzy Logi Control. Gerard S hram, Uzay Kaymak and Henk B. Ver-
bruggen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

RCAM part
14 The RCAM Design Challenge Problem Des ription. Paul Lambre hts,
Samir Bennani, Gertjan Looye and Dieter Moormann . . . . . . . . . . . . . . 149

15 The Classi al Control Approa h. Jim E. Gautrey . . . . . . . . . . . . . . . . . . . . 180

16 Multi-Obje tive Parameter Synthesis (MOPS). Hans-Dieter Joos . . . . . 199

17 An Eigenstru ture Assignment Approa h (1). Lester Faleiro and Roger

Pratt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218

v
18 An Eigenstru ture Assignment Approa h (2). Jess M. de la Cruz, Pablo
Ruiprez and Joaqun Aranda . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238

19 A Modal Multi-Model Approa h. Carsten Dll, Jean-Franois Magni and

Yann Le Gorre . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258

20 The Lyapunov Approa h. Jamal Daafouz, Denis Arzelier, Germain Gar ia

and Ja ques Bernussou . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278

21 An H1 Approa h. Mark R. Tu ker and Daniel J. Walker . . . . . . . . . . . . 300

22 -Synthesis Approa h (1).

A Samir Bennani and Gertjan Looye . . . . . . 321

23 A -Synthesis Approa h (2). Jan S huring and Rob M.P. Goverde . . . 341

24 Autopilot Design based on the Model Following Control Approa h. Holger

Duda, Gerhard Bouwer, J.-Mi hael Baus hat and Klaus-Uwe Hahn 360

25 Flight Management using Predi tive Control. Mihai Huzmezan and Jan M.
Ma iejowski . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379

26 A Fuzzy Control Approa h. Gerard S hram and Henk B. Verbruggen . 398

HIRM part
27 The HIRM Design Challenge Problem Des ription. Ewan Muir . . . . . . 421

28 Design via LQ Methods. Fran es o Amato, Massimiliano Mattei, Stefano

S ala and Leopoldo Verde . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446

29 The H1 Loop Shaping Approa h. George Papageorgiou, Keith Glover and

Ri k A. Hyde . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 466

30 Design of Stability Augmentation System using -Synthesis. Karin Sthl

Gunnarsson . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 486

31 Design of a Robust, S heduled Controller using -Synthesis. Johan An-

thonie Markerink . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 505

32 Nonlinear Dynami Inversion and LQ Te hniques. Batri e Es ande . . 525

33 The Robust Inverse Dynami s Estimation Approa h. Ewan Muir . . . . . 543

Con luding part

34 The Industrial View. Chris Fielding and Robert Lu kner . . . . . . . . . . . . 569

35 An Other View of the Design Challenge A hievements. Georg Grbel 605

36 Con luding Remarks. Samir Bennani, Jean-Franois Magni and Jan Ter-
louw . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 612

Appendix
A Used Nomen lature. Anders Helmersson and Karin Sthl Gunnarsson 614

Bibliography

vi
Author Index

Fran es o Amato . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33, 446

Joaqun Aranda . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
Denis Arzelier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42, 278
J.-Mi hael Baus hat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116, 360
Samir Bennani . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81, 149, 421, 321, 612
Ja ques Bernussou . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42, 278
Gerhard Bouwer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116, 360
Jess M. de la Cruz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22, 238
Jamal Daafouz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42, 278
Binh Dang Vu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
Carsten Dll . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
Holger Duda . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116, 360
Batri e Es ande . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 525
Pierre Fabre . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
Lester Faleiro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22, 218
Chris Fielding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1, 569
Germain Gar ia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42, 278
Jim E. Gautrey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
Keith Glover . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64, 466
Rob M.P. Goverde . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
Georg Grbel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13, 605
Klaus-Uwe Hahn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116, 360
Anders Helmersson . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149, 614
Mihai Huzmezan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125, 379
Ri k A. Hyde . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 421, 466
Jonathan Irving . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 421
Joseph Irvoas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
Hans-Dieter Joos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13, 199
Uzay Kaymak . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
Paul Lambre hts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149, 421
Tony Lambre gts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
Yann Le Gorre . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
Gertjan Looye . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81, 149, 321
Robert Lu kner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 569
Jan Ma iejowski . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125, 379
Jean-Franois Magni . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22, 149, 258, 612
Johan Anthonie Markerink . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 505
Alberto Martnez . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
Massimiliano Mattei . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33, 446
Philippe Mnard . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
Dieter Moormann . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149, 421
Ewan Muir . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112, 421, 543

vii
George Papageorgiou . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64, 466
Ian Postlethwaite . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
Roger Pratt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
Pablo Ruiprez . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
Stefano S ala . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22, 33, 149, 421, 446
Carsten S herer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
Gerard S hram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135, 398
Jan S huring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341, 421
Phillip Sheen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
Alex Smerlas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
Karin Sthl Gunnarsson . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 421, 486, 614
Jan Terlouw . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1, 149, 421, 612
Mark R. Tu ker . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52, 300
Hans van der Vaart . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
Henk B. Verbruggen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135, 398
Leopoldo Verde . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446
Daniel J. Walker . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52, 300

viii
1. Introdu tion

Jan Terlouw and Chris Fielding

1 2

1.1 The Importan e of Advan ed Control

Design Methods for the European Air raft
Industry
European manufa turers of military and ivil air raft have rea hed a high level
of expertise in designing ight ontrol laws, to a point that they an solve
virtually any realisti hallenge that might be foreseen in the near future. This
apability is a result of the lessons learned by generations of engineers who have
extended and passed on their skills, always driven by the ultimate requirement
- that one day their ight ontrol system (FCS) had to y. However, the
large time and eort spent to solve all problems en ountered during the design
pro ess poses the question whether improvements are possible.
As the s ienti ommunity sometimes laims to have invented new methods
to improve urrent ways of working, there is a natural interest from industry in
what the resear hers have to oer. On the other hand, s ientists are interested
in realisti appli ations to justify their work and to test new on epts.
It is lear that there is a strong in entive for both worlds to work together,
but a tually a hieving it an be di ult. S ientists like to develop methods
whi h have general appli ability, and this is parti ularly true for ontrol theo-
reti ians. On the other hand, users of design methods are, from a professional
point of view, mainly interested in dedi ated methods that solve their parti -
ular problems. The result is that many new ideas never really break through,
be ause they are simply not spe ialised and elaborated enough, or be ause
there is not enough liaison between the s ienti and industrial worlds.
There are now a large number of ontroller design methods that have been
developed over the past twenty-ve years (some have earlier origins). In this
book twelve of them are treated:

 Multi-Obje tive Parameter Synthesis

 Eigenstru ture Assignment

 Linear Quadrati Optimal Control

1
National Aerospa e Laboratory NLR, Flight Me hani s Department, Anthony Fokker-
weg 2, 1059 CM Amsterdam, The Netherlands.
2
British Aerospa e Military Air raft, Aerodynami s Department, Warton Aerodrome,
Preston PR4 1AX, UK

1
  Lyapunov Te hniques
 H1 Mixed Sensitivity
 H1 Loop Shaping
 -Synthesis
 Nonlinear Dynami Inversion
 Robust Inverse Dynami s Estimation
 Model Following
 Predi tive Control
 Fuzzy Logi Control
These methods have many dierent features. A ommon feature is that ea h of
them is developed to a hieve advantages over lassi al te hniques. The laimed
benets range from enhan ed performan e, resulting from multi-input multi-
output ontrollers, to improved e ien y and simpli ation of the design pro-
ess. At the same time, the most important and obvious di ulty in adopting
any new method is the la k of experien e of its use in pra ti e.
This book is an attempt to redu e the gap between theory and prati e,
with respe t to appli ation of modern ontrol design te hniques. It deals with
ight ontrol of rigid body ivil and military air raft. The twelve te hniques
mentioned above will be demonstrated on the basis of two ben hmark problems
[145, 177. But rst, some general remarks will be made about ight ontrol
laws as a part of FCS design.

Flight ontrol laws design

The main fun tion of the ight ontrol system (FCS) of an air raft is to on-
tribute to its safe and e onomi operation, su h that the intended ight missions
an be a omplished and unexpe ted events an be handled. The heart of a
modern FCS onsists of the following omponents, arranged in a logi al way
to benet from the prin iple of feedba k: sensors provide a ight ontrol om-
puter (FCC) information on air data, inertial data and o kpit data; an FCC
in whi h ight ontrol laws are implemented to determine the ommands for
the a tuation systems of the air raft ontrol surfa es and throttles for engines
demands. For air raft, feedba k ontrol is used to provide tight pilot ommand
tra king, to attenuate external disturban es su h as gusts and turbulen e and
to provide robustness against modelling errors.
In the early days of ight, safety was the main on ern for FCS design-
ers. Pilots needed signi ant eort to maintain some ight onditions under
all ir umstan es. Today, safety is even more important, be ause many more
people are transported, higher osts are involved in establishing safety, and
the reputation of airlines and air raft manufa turers is paramount, in an in-
reasingly ompetitive market. Fly by Wire allows the pilot to ontrol the

2
air raft states, as an alternative to the onventional dire t ontrol of the engines
and ontrol surfa es. It gives new opportunities to in rease the overall level
of safety through the exibility oered by the ontrol laws [78. For example,
error-tolerant ontrol laws provide ight envelope prote tion, and help the pilot
to re over from unusual attitudes and su essfully a hieve riti al manoeuvres.
The use of modern FCS an be bene ial from an e onomi point of view.
For ertain types of air raft, fuel onsumption an be redu ed by allowing re-
laxed stati stability, ountera ted by the appli ation of a tive ontrol. Another
advantage related to fuel onsumption is that for large air raft the weight of Fly
by Wire systems is smaller than that of onventional systems. Furthermore,
the so- alled family on ept an be introdu ed. Flying dierent air raft an
be made almost the same for pilots, by making appropriate adjustments in
the ight ontrol laws. As a result, dierent air raft feel almost the same,
therefore helping to redu e pilot training osts.
Most importantly, modern FCS have ontributed to improved dynami al
behaviour. Certain military air raft annot be own without a stability aug-
mentation system. The open loop instability, whi h is related to agility of
the air raft, is utilised to obtain better performan e and manoeuvrability of
the losed-loop system. For ivil air raft, performan e an be in reased by
appli ation of a tive systems, for example to provide gust suppression and
auto-trimming, in order to a hieve improved ride quality.
The performan e benets a hieved, have the penalty of tremendous osts
involved in the development of an advan ed FCS. In the past, the pilot sti k was
typi ally onne ted with rods or ables to the ontrol surfa es. Sin e then, the
in reased safety, and e onomi al and performan e demands have for ed air raft
manufa turers to extend FCS to a high level of omplexity. The danger exists
that the e onomi al benets des ribed above are nullied by higher design and
maintenan e osts, while omplexity an potentially have a negative ee t on
safety.
The large number of fun tions and requirements have in reased the number
of spe ialists areas needed for the FCS design pro ess. This makes the work
hallenging from a te hni al and management point of view. People who are
responsible for mode logi , redundan y design, software and hardware develop-
ment, design integration and erti ation have to work losely together. In the
overall pro ess, ontrol laws designers assume a modest, but entral position.
They have knowledge of ight me hani s, ontrol theory, handling qualities,
airframe stru ture and FCS hardware. Their task is inuen ed by the de-
sign requirements, the ight envelope, the air raft onguration omplexity,
the stores arriage and weight distribution, the required autopilot modes, the
air raft stability (or instability) levels and the aerodynami nonlinearity.
The work of an industrial ight ontrol laws designer who uses lassi al
design te hniques (see Chapter 15) may onsist of the following simplied se-
quen e of a tivities. The rst step is to derive a nonlinear dynami model of
the air raft to be ontrolled. Getting familiar with the dynami al behaviour
by means of trimming, stability and ontrol analysis and nonlinear simulations
(for stable air raft) and understanding the inuen es of the modelling assump-

3
tions is most important at this stage. Linearisation and linear simulation of the
model is also performed. The next step is to dene the ontroller ar hite ture
and to make a rst design whi h in ludes gain s heduling to over the air raft's
ight envelope. Implementation of the ontrol law in the nonlinear model, for
o-line and piloted simulation, is arried out next. This pro ess might be re-
peated to optimize the design. In the design pro ess, nonlinearities and model
un ertainties are important issues to understand and deserve mu h attention
if a robust design is to be a hieved.

Robustness of ight ontrol systems

Robustness investigation deals with the dis repan y between models and re-
ality. It is basi ally on erned with whether a ontrolled system will work
satifa torily under the ir umstan es it will meet in pra ti e. FCS designers
have always used models to in rease their knowledge about ight ontrol, and
have been invloved in robustness investigations in some form, sin e the very
beginning of ight.
At various stages of the ight ontrol laws design pro ess, model un ertainty
an be introdu ed, for example, when linearised versions of omplex models are
derived. In this ase, the term un ertainty is a tualy a misnomer, be ause the
deviation between linear and nonlinear behaviour an be quantied. The same
is true if known variations in, for example, the position of the entre of gravity
or a time-delay in the system, are negle ted. Depending on the design te hnique
used, it may be ne essary to make su h modelling assumptions (temporarily)
in order to obtain a model whi h is suitable for ontroller design.
Model un ertainty an also be introdu ed unintentionally due to modelling
errors, unknown hara teristi s of the air raft in relation to the environment,
or ina urate information about the signals owing through the system. For
example, the pre ise value of aerodynami stability derivatives and air data
may not exist.
A feature of several ontrol design te hniques des ribed in this book is that
they deal systemati ally and sometimes expli itly with robustness. Introdu ing
these systemati s into the design y le may enhan e FCS design in terms of the
ee ien y of the pro ess and the performan e of the resulting ontrol systems.

Potential ontribution of modern design te hniques

It seems that the European aeronauti al industry is not in the rst pla e in-
terested in modern te hniques purely to a hieve better air raft performan e.
In fa t almost any te hnique, modern or lassi al, when used to solve realis-
ti problems with enough knowledge of the method itself, with the ne essary
tools available, and based on a thorough knowledge of ight me hani s, will
eventually lead to the desired results. The real interest is in the systemati ap-
proa hes behind new methods, be ause this an simplify the design y le and
make it more transparent. Global ompetition for es air raft manufa turers to
ontinuously improve the e ien y of their engineering a tivities. If it an be

4
demonstrated that advan ed design te hniques lead to a design y le with bet-
ter tra ability of design de isions and simpliation of the overall pro ess, the
han e that modern ontrol te hniques will be used by industry will in rease.
The omplexity of the design task and the related investment made in the
past in human and non-human apital, explain the areful attitude from some
air raft manufa turers to repla e their well-established lassi al te hniques.
Moreover, lassi al te hniques have desirable features, for example the visi-
bility of the resulting ontroller. At the lowest level of detail of the ontrol
law, the fun tion of every gain and dynami element an be easily understood,
whi h makes designs easy to modify and a ept. On the other hand, the visi-
bility after integration of subsystems is partly lost at a higher level. Another
advantage is that gain and phase margins are open-loop measures with a lear
link to robustness. This makes them very useful for synthesis.
Even though it is true that superiority is often related to simpli ity and
transparan y, whi h are typi al features of lassi al ontrol te hniques, the
aeronauti al industry a knowledges some disadvantages as well. Due to his-
tori reasons, the lassi al approa h in whi h ea h mode and ight ondition
is treated as a separate problem has led to mode proliferation and the need
for omplex algorithms. To avoid fun tional integration at the end of the FCS
design, whi h is too late, an all en ompassing and onsistent design strategy is
ne essary. Throughout the design pro ess a systems approa h strategy should
be applied, supported by good requirements, design tools and design models.
Appli ation of advan ed te hniques promises a signi ant redu tion of design
time be ause it would remove the time- onsuming lassi al one-loop-at-a-time
approa h and redu e the number of design points for whi h a ontroller has to
be designed.

1.2 GARTEUR A tion Group on Robust Flight

Control
In O tober 1994, GARTEUR
3 A tion Group FM(AG08) was established. For
the twenty-three member organisations of this group from seven European
ountries, GARTEUR proved to be an organisation oering ex ellent ondi-
tions and support for arrying out basi , pre ompetitive resear h. GARTEUR
unites resear h establishments, the aeronauti al industry and universities in
A tion Groups. In FM(AG08) the following organisations parti ipated:

 Resear h Establishments

 Centro Italiano Ri er he Aerospaziali (CIRA, Italy, Capua)

3
The Group for Aeronauti al Resear h and Te hnology in EuRope (GARTEUR) was
formed in 1973 and has as member ountries: Fran e, Germany, The Netherlands, Spain,
Sweden and the United Kingdom. A ording to its Memorandum of Understanding, the
mission of GARTEUR is to mobilize, for the mutual benet of the GARTEUR member
ountries, their s ienti and te hni al skills, human resour es and fa ilities in the eld of
aeronauti al resear h and te hnology. More information about GARTEUR an be found in
the GARTEUR Guide [4.

5
  Deuts he Fors hungsanstalt fr Luft- und Raumfahrt (DLR, Ger-
many, Oberpfaenhofen)

 Defen e Resear h Agen y (DRA, United Kingdom, Bedford)

 Instituto Na ional de T ni a Aeroespa ial (INTA, Spain, Madrid)

 Laboratoire d'Automatique et d'Analyse des Systmes (LAAS, Fran e,

Toulouse)

 National Aerospa e Laboratory (NLR, The Netherlands, Amster-

dam)

 O e National d'Etudes et de Re her hes Arospatiales

 CERT-ONERA, Fran e, Toulouse

 ONERA-Salon, Fran e, Salon de Proven e

 Industry

 Alenia Aeronauti a (ALN, Italy, Turin)

 Avro International Aerospa e (AVRO, United Kingdom, Woodford)

 British Aerospa e, Dynami s (BAe-D, United Kingdom, Filton)

 British Aerospa e, Military Air raft (BAe-MA, United Kingdom,

Warton)

 Cambridge Control (CCL, United Kingdom, Cambridge)

 Daimler Benz Aerospa e Airbus (DASA, Germany, Hamburg)

 Fokker Air raft Company (FAC, The Netherlands, Amsterdam)

 Saab Military Air raft (SMA, Sweden, Linkping)

 Universities

 Craneld University (CUN, United Kingdom, Craneld)

 Delft University of Te hnology (DUT, The Netherlands, Delft)

 Linkping University (LiTH, Sweden, Linkping)

 Loughborough University (LUT, United Kingdom, Loughborough)

 University of Cambridge (UCAM, United Kingdom, Cambridge)

 University of Lei ester (ULES, United Kingdom, Lei ester)

 Universit di Napoli "Fediri o II" (UNAP, Italy, Naples)

 Universidad Na ional de Edu a in a Distan ia (UNED, Spain,

Madrid)

The A tion Group was haired by NLR (Terlouw); CERT-ONERA (Magni)

provided the vi e- hairman. In total eight meetings were held in Amster-
dam, Madrid, Cambridge, Capua, Toulouse, Linkping, Oberpfaenhofen and
(again) Amsterdam, whi h gave an extra ultural dimension to the proje t.
In view of the longer term obje tive of ontributing to e ien y improvement
of the ight ontrol laws design pro ess, it was de ided to follow three main
streams.

6
Design Challenge
The rst stream was the Design Challenge des ribed in this book. Before the
start of the A tion Group it was on eived that a thorough demonstration
of modern design te hniques, applied to genuine ight ontrol problems, was
required in order to get the desired feedba k from industry. The aim was to
present the state-of-the-art with respe t to modern (robust) ontrol in su h a
way that industry ould relate to it. At the same time it was the intention to
larify what is needed for a design method to be a epted by an industrial design
o e. To a hieve this, people from industry were asked to give inputs for two
ben hmark problems, whi h were subsequently developed by people from the
resear h establishments and universities. The rst one, the RCAM (Resear h
Civil Air raft Model) problem [145, is based on the automati landing of a
large, modern argo air raft. The se ond, the HIRM (High In iden e Resear h
Model) problem [177, onsiders the ontrol of a military air raft a ross a wide
design envelope.
Both ben hmarks are based on six degrees of freedom mathemati al air raft
dynami s models, dened in Matlab/Simulink [121, 240. They in lude aero-
dynami , engine, atmosphere and gravity models. In addition, a tuator and
sensor hara teristi s are taken into a ount, together with models for wind, at-
mospheri turbulen e and windshear. An extensive set of design requirements is
given, whi h an be tested with software for frequen y and time domain eval-
uations. A standard nomen lature [237 and a standard report lay-out were
dened at an early stage, to avoid unne essary problems later on. In order to
make the ben hmarks more realisti , parameter variations (time-delay, mass
and entre of gravity variations for RCAM; variations in aerodynmi deriva-
tives and measurement errors for HIRM) were dened. Furthermore, some
hardware implementation issues are onsidered. This puts the ben hmarks
into the ategory of robust ight ontrol problems.
At the start of the proje t it was de ided to limit the s ope of the demon-
stration of the te hniques to design and omputer simulations. Validation of
the most promising ontrol laws and design te hniques might possibly be per-
formed in a follow-on proje t, in whi h the use of a ight simulator and a ying
testbed is re ommended.
The Design Challenge was not aimed at giving the answer to the question
whi h method is best?, but rather to show, step by step, how modern ontrol
an be applied. The design teams were asked to highlight four main points:

1. The eort to learn, to implement and to apply the method.

2. The appli ability of the design method to ight ontrol laws design.

3. The omplexity of the resulting ontroller, its implementation and erti-

 ation issues.

4. The robustness and performan e of the designed ontroller.

A large group of ontrol engineers primarily from the European aeronauti al

7
industry has evaluated the proposed designs. This book is a summary of the
results of the Design Challenge.

Computer-Aided Control System Design

The se ond stream of a tivities addressed the development of a framework for
omputer-aided ontrol system design. Several industrial members of GAR-
TEUR A tion Group FM(AG08) indi ated a need for omputer support of the
design pro ess and data. A prototype was developed by NLR, based on the
ommer ial software produ ts Matlab/Simulink, SiFrame and Informix. The
prototype oers fa ilities for design pro ess denition and exe ution, in luding
tool integration and a entral data repository. Most important is the automati
onsisten y ontrol of all (versions of ) design information in the framework.
The  lassi al design pro ess of Craneld University, des ribed in Chapter 15,
is implemented in the prototype, whi h was evaluated by several FM(AG08)
organisations. The results of this eort are des ribed seperately in [224.

Robust Flight Control Tutorial and Literature Survey

Database
The third stream aimed at making available a literature overview of ontrol
te hniques related to ight ontrol appli ations and at providing a tutorial
do ument on advan ed ontrol te hniques.
CIRA has established a Robust Flight Control Literature Survey Data Base,
whi h an be a essed via the Internet [206. From it, referen es and summaries
of arti les on robust ight ontrol an be obtained. The aim of the database
is to assist designers in lassifying their ontrol problems a ording to similar
problems already treated in the literature. As su h, it an help the designer to
identify the most onvenient te hnique to be used.
INTA has written a do ument [53 in whi h tutorials of all design te hniques
that are des ribed in this book and several others are presented in detail.

1.3 Outline of the book

The editors would like to point out that this book is the result of a group
a tivity. With respe t to the ontents, it was onsidered to be important that
as many FM(AG08) organisations as possible would get the opportunity to
present their results, in order to over a wide variaty of design approa hes.
The ontributions were not sele ted by the editors.
The book onsists of four parts. Part I ontains tutorials of all methods
that have been applied on either RCAM or HIRM or on both problems.
Parts II and III over the RCAM and HIRM ben hmark denions and the
proposed designs. With a few ex eptions, ea h design hapter has basi ally
the same stru ture. The designs are based on the twelve methods given in
se tion 1.2. Furthermore, one RCAM design is entirely based on lassi al te h-
niques.

8
 In Part IV, three dierent views on the Design Challenge are given. Chap-
ter 34 presents a view from industry. A questionnaire was designed by British
Aerospa e and DASA to aid evaluators in their assessment of the Design Chal-
lenge entries. Chapter 35 dis usses the Design Challenge results from the s i-
enti resear her's point of view. An obje tive measure of stability robustness,
namely the stru tured singular value, is given for ten RCAM designs. Finally,
Chapter 36 ontains some on luding remarks of the editors.
One of the onditions whi h made the Design Challenge possible was the fa t
that all teams have used the same nomen lature, whi h is given in Appendix A.

A knowledgements
Most of the work needed for writing this book was funded by the parti ipating
organisations of GARTEUR A tion Group FM(AG08). These organisations,
whi h are listed in se tion 1.2, are given thanks for their onden e in the
group and their full support until the end of the proje t. In some ases national
agen ies and other resear h funding bodies have given additional nan ial help,
notably the Netherlands Agen y for Aerospa e Programs (NIVR). Without
their support the Design Challenge would not have been possible.
FM(AG08) also wishes to express its gratitude to Arospatiale and DRA
for making available the models on whi h the RCAM and HIRM ben hmark
denitions are based.
Another word of thanks is to the GARTEUR organisation, in parti ular
the Flight Me hani s Group of Responsables and the Exe utive Committee,
for making the publi ation of this book possible. The head of the NLR Flight
Division, Jan van Doorn, who has a ted as the GARTEUR Monitoring Re-
sponsable of FM(AG08), has given essential ontributions behind the s enes.
He was an indispensible link between the A tion Group and the GARTEUR
organisation.
The editors are grateful to Chris Fielding, Derek Laidlaw, Jim Gautrey,
Lester Faleiro, Danil Walker and Jonathan Irving for he king most hapters
on the use of English and proposing many suggestions for improvements.
Not all results of GARTEUR A tion Group FM(AG08) related to the Design
Challenge ould be presented in this book. Several design teams joined in
at a later stage or there were other reasons why their designs ould not be
in luded. In this respe t Alex Smerlas (Univ. of Lei ester) [217, Aldo Tonon
(ALN), Jrgen A kermann (DLR), Angel Perez de Madrid (UNED) and their
olleagues are a knowledged for their valuable ontributions.
This book will be presented at a GARTEUR Spe ialists' Workshop on Ro-
bust Flight Control (CERT-ONERA, Toulouse, Fran e, April 14-15, 1997).
Spe ial thanks is given to CERT-ONERA for organising and hosting this work-
shop.

9
10
 Part I

Tutorial part

11
2. Multi-Obje tive Parameter Synthesis
(MOPS)

Georg Grbel1 and Hans-Dieter Joos 1

2.1 Theoreti al Aspe ts

2.1.1 Global Goals
Any ontrol law is parametrized in some way. For example, in a P-I-D ontrol
stru ture with additional shaping lters there are the P-I-D gains and the lter
parameters whi h are to be tuned for losed-loop performan e. Similarly, an
LQR state- or output feedba k ontrol law is parametrized by the state- and
ontrol eort weights Q and R, an eigenstru ture state feedba k ontrol law is
parametrized by the eigenvalues and some eigenstru ture parameters and an
H1 ontrol law is parametrized by its input/output weighting lter parameters.
Control parameter tuning for a best possible robust performan e is a time-
onsuming task if performed manually. This is due to the multitude of dif-
ferent (nonlinear) design spe i ations whi h have to be dealt with. This is
also true if one uses advan ed analyti al ontrol synthesis te hniques su h as
 synthesis.
Multi-obje tive parameter synthesis (MOPS) is a general te hnique whi h
omplements a hosen ontrol law synthesis te hnique. Having hosen an
appli ation-spe i ontrol law stru ture with parametrization, or having ho-
sen a general ontrol synthesis te hnique with its analyti ally given parame-
terization, the free design parameters (e.g. the LQR-weights) are omputed
by a min-max parameter optimization set up. The designer formulates this
set up by spe ifying the design goals as a set of well dened omputational
riteria, whi h an be a fun tion of stability parameters (e.g. eigenvalues),
and time- and frequen y response hara teristi s (e.g. step-response overshoot
and settling time, ontrol rates, bandwidth, stability margins et ). By this
multi- riteria formulation all the various oni ting design goals are taken are
of individually, but are ompromised on urrently by a weighted min-max pa-
rameter optimization. In parti ular, robust- ontrol requirements with respe t
to variations in stru tured parameter sets and operating onditions an be
taken are of by a multi-model formulation whi h en ompasses the worst- ase
design onditions.

1
DLR German Aerospa e Resear h Establishment, Institute for Roboti s and System
Dynami s, Control Design Engineering Group (Prof. G. Grbel), D-82234 Wessling
E-mail: dieter.joosdlr.de

13
 For oni ting design riteria the te hnique provides a best-possible pareto-
optimal ontrol tuning. Sin e the multi- riteria in ludes performan e measures
whi h are dire tly used as design drivers, they provide dire t quantitative infor-
mation about the design oni ts and performan e onstraints. This yields all
the ne essary information on how to improve the balan e of a design trade-
o within a given ontroller stru ture or a hosen ontrol synthesis method.
The method allows simple (linear) ontrollers to be optimized based on om-
plex (nonlinear) system evaluation models, thereby linking together the steps
of ontrol design and of (nonlinear) design assessment. Our experien e shows
that with the same engineering eort, a dedi ated ontrol performan e improve-
ment of about 30% an be a hieved by numeri al multi-obje tive optimization
as ompared to manual design parameter tuning in a sequential manner.

2.1.2 System Model Des ription

Both linear and nonlinear design models an be taken into a ount. In a multi-
model approa h linear models together with nonlinear models an be used
simultaneously. There is no restri tion on the representation of system distur-
ban es.

Robustness against stru tured parameter deviations or, for example, sensor
failures is a hieved by applying a ommon ontroller to a set of xed worst- ase
parameter models. This model set hara terizes the worst dynami s deviations
within the range of operation, e.g. the ight envelope, or a part of it. For ea h
su h model the appropriate set of riteria has to be spe ied. Hen e, the multi-
model problem is transformed into a multi- riteria optimization problem.

In general, there exists no theory that guarantees stability or performan e

robustness a ross the range of operation, if only a nite number of operating
points is onsidered simultaneously. It depends on the physi al properties of the
system to be designed, whether runaways an exist. If they exist, they have to
be added to the set of worst- ase operating points and treated simultaneously
by the multi-model approa h. Worst- ase operating points an be omputed
by a dual appli ation of the parameter optimization approa h: Find those
parameter ombinations within a given un ertainty domain whi h yield the
worst performan e for the hosen multi- riteria set up [20.

Of ourse, robustness of the ontroller about an operating point an be

enfor ed in the multi-obje tive approa h by adding suitable robustness riteria
(e.g. gain/phase margins) to the set of otherwise spe ied performan e riteria.

2.1.3 Controller Stru ture

Both linear ontrollers and nonlinear ontrollers (e.g. fuzzy ontrollers or adap-
tive ontrollers) an be used. If a spe i analyti al synthesis te hnique is ap-
plied within this framework, the ontroller stru ture is bound to this synthesis
te hnique.

14
2.1.4 Design Spe i ations
Ea h design obje tive may be mathemati ally des ribed by a well-dened ri-
terion i whose value is the smaller, the better the obje tive is a omplished.
Examples taken from the RCAM design hallenge spe i ations are:

- Steady state error, settling time and rise time for demanded steady state
value ys :
Z tend
= (y(t) ys )2 dt
t1

- Overshoot over demanded steady state value ys :

= max
t
(y(t)=ys )

- Rise time dened as the time the unit step response y(t) takes from
y(t1 ) = 0:10 to y(t2 ) = 0:90:
= t2 t1
- Minimal damping of the eigenvalues evi :
= 1 min
i
( Re(evi )=jevi j) :

In order to get smooth riteria as a fun tion of the tuning parameters, the
min- or max-fun tions involved are smoothed by an exponential approximation;
see also se tion 2.5. The above eigenvalue riterion minimal damping is re-
formulated in su h a manner that better damping results in a smaller riterion
value.

2.1.5 Analysis Information

To evaluate the hosen design riteria i , the respe tive analysis omputations
have to be performed. This usually requires eigenvalue omputations, time
response simulations and frequen y response omputations. This analysis in-
formation is used to judge the quality of a design in addition to the riteria
values whi h quantify the balan e of the a hieved optimum, and whi h provide
further ontroller synthesis information (see se tion 2.1.6).

2.1.6 Controller Synthesis Information

To ea h riterion, i an upper-bound demand value or driver value di is
dened by the designer. Then the tuning parameters T = [Tk are omputed
by solving the min-max parameter optimization problem

min max f i =di g

T i

15
subje t to performan e and tuning onstraints:

gj (T )  0; Tkmin  Tk  Tkmax:
This is the MOPS synthesis formula. By iterating the demand values di
as a fun tion of the a hieved riteria values i 1, the resulting ompromise
trade-o solution an be driven in a desired dire tion.

2.1.7 Pra ti al Implementation Aspe ts

The overall omputing time for the synthesis mainly depends on the time for
riteria evaluations. Hen e fast algorithms and software implementations [101
are required for the orresponding analysis omputations. It is good pra ti e
to use  heap riteria where possible. The number of riteria evaluations also
depends on the number of models used in a multi-model set-up. Therefore it
is also bene ial to minimize the number of models by a areful sele tion of
worst- ase operating points or parameter deviations.

2.1.8 Relation with other Methods

Multi-obje tive parameter synthesis loses the parametri design loop with
modelling-, synthesis- and analysis methods a ording to Figure 2.1.

synthesis plant
model model

D T P controller closed-loop M simulation/

synthesis
model model analysis

C performance/cost I
criteria

Figure 2.1: Design loop losed by multi-obje tive parameter synthesis.

It serves to automate ontrol tuning to given performan e spe i ations. It

is neutral as far as the design steps modelling, synthesis and analysis are
on erned.

2.2 Example of Appli ation

The approa h has been applied for robust ight ontrol [138, 102, 100, a tive
antenna-beam ontrol [19, PWM-satellite attitude ontrol [98, maglev vehi le

16
 ontrol [190, (semi-)a tive ar suspension and air raft landing gear ontrol [81,
209, 254, robot ontrol [153, and others. An example appli ation is the MOPS-
solution for the RCAM design hallenge [130. There, for the longitudinal
ontrol, LQR PI-output feedba k is used, whereas for lateral ontrol a lassi al
ontrol stru tue [35 is used, thereby demonstrating the appli ation for two
dierent ontroller stru tures. A nonlinear worst- ase plant analysis, also using
MOPS, was performed to he k robustness within the multi-model set-up.

2.3 Computational Aspe ts

The method requires the set-up of a omputation loop a ording to Fig. 2.1
and the availability of a suitable min-max parameter optimization software.
For an engineering-e ient appli ation of this te hnique it is very bene ial to
have a software framework whi h supports intera tive modular problem setup
and demand spe i ation as well as automated performan e evaluation (su h
as ANDECS_MOPS [99).
Multi-model/multi-obje tive performan e evaluation an be fun tionally
parallelized, e.g. by using the PVM (Parallel Virtual Ma hine) lient-server
network on ept. Thereby the omputation time an be redu ed.
The multi-obje tive optimization problem an be solved by any nonlinear
programming tool, sin e minimizing a set of riteria an be transformed into a
onventional s alar nonlinear programming problem; see 2.5.
Using appli ation-spe i engineering riteria in pra ti e, typi ally leads to
non onvex optimization problems. Thus lo al minima may exist. However, a
lo al minimum solution is also a lo al best-possible pareto-optimal solution. If
su h a solution is not satisfa tory, other solutions an be found by hanging the
demand values, or by hanging the starting values for the tuning parameters.
To avoid lo al minima, a global optimizer has to be used whi h may have the
disadvantage of rather long omputing times.

2.4 Comparative Study

Multi-obje tive parameter synthesis allows full exploitation of a given on-
troller stru ture, as a fun tion of the ontroller parametrization. In parti -
ular, it allows the exploitation of the a hievable trade-os between ontrol
performan e and required ontrol eort. This is possible in a most detailed,
appli ation-spe i way and hen e, no matter what ontroller stru ture or on-
troller synthesis method is used, this te hnique, in prin iple, always yields the
best possible performan e in the hosen ontext.
A potential benet of multi-obje tive tuning the design parameters of an
analyti synthesis method (e.g. LQR, eigenstru ture synthesis, et .) instead
of dire tly tuning the parameters of a given ontroller stru ture (i.e. state- or
dynami output feedba k), is primarily that usually a smaller number of pa-
rameters is to be tuned. This parti ularly holds for multi-input/multi-output

17
systems. Also, built-in performan e and robustness features of the hosen syn-
thesis method are automati ally guaranteed. On the other hand, dire tly tun-
ing the parameters of a spe ied ontroller stru ture allows the designer to use
appli ation-proven ontroller stru tures for whi h no analyti synthesis te h-
niques exist, and it allows him to extend and to adapt su h stru tures during
the design pro ess.
If an appropriate software framework is available whi h provides a pre-
dened omputation loop and a set of standard riteria to hoose from (e.g.
ANDECS), the level of required training is moderate. In this ase, no spe i
mathemati al theory is required. Design spe i ations are expli itely formu-
lated in their most natural mathemati al form and a trans ription of design
spe i ations into a synthesis-spe i weighting form is not required.
In omplex design-de ision problems with, say, more than 5 riteria to be
handled simultaneously, an integrated data system is mandatory, to keep tra k
of the de ision iterations made during the design pro ess (this holds for any
design-iteration logi ). The ANDECS software provides su h an integrated data
system, whi h is spe i ally-designed for multi-obje tive/multi-model design
iterations.

2.5 Mathemati al Appendix

2.5.1 Preferen e order, goal oni ts and satisfa tory om-
promise sets for riteria ve tors
The main advantage of a multi-obje tive design is the possibility to formulate an
individual riterion for ea h spe ied demand, while treating all riteria during
optimization simultaneously. Here, some terms are laried by introdu ing the
related on epts [128:

- better solution in the ontext of a preferen e order for ve tor-valued

riteria,

- goal oni t and  ompromise in the ontext of pareto-optimality and

- satisfa tory ompromise in the ontext of demand level.

(i) The individual riteria i (T ) are ombined to give a riteria ve tor (T ).

The following preferen e order allows one to partially ompare su h ve -
tors:

A set of tuning parameters T1 is said to be better than T 2 , if

the orresponding riteria ve tor (T 1) is smaller than (T 2),
where smaller means

(T 1) < (T 2 ) , i (T 1) < i (T 2 ) for all i

18
 Smaller equal is dened as

(T 1)  (T 2 ) , i (T 1)  i (T 2 ) for all i
1 2
and i (T ) < i (T ) for at least one i.

(ii) Trying to improve several riteria simultaneously normally leads to a goal

oni t in the sense that no riterion an be improved further without
worsening another one. More pre isely:

A set of tuning parameters T  is alled a ompromise solution,

or pareto-optimal solution, if
there is no T with (T ) < (T  ).
(iii) Usually, ompromise solutions are not unique. There exists a whole set
of pareto-optimal solutions and it is up to the design engineer to de ide
what trade-o is a best satisfa tory ompromise in his design ontext.
The term satisfa tory an be made more pre ise by introdu ing the
demand level d referring to a riteria ve tor :
A set of tuning parameters T  forms a satisfa tory ompromise,
if T  belongs to the set of ompromise solutions and if

(T )  d ;
where in d the demands of the designer are quantied. The
ve tor d is alled the demand level.

Fig. 2.2 illustrates the above denitions for the ase of 2 riteria 1 and
2 .
Assume that fT g denotes the set of all feasible tuning parameters T and
that (fT g) is the orresponding value set. The thi k border part of (fT g)
in Fig. 2.2 is the set of ompromise solutions and Cs marks the subset of a
satisfa tory ompromise. Note, that all solutions with riteria values smaller
than the demand level d are satisfa tory solutions.

2.5.2 Finding a satisfa tory ompromise set by means of

min-max optimization
A parti ular, satisfa tory ompromise an be found by means of parameter
optimization.

From the riteria ve tor (T ) and the demand level d one an form
a s alar fun tion

 = max f i (T )=di g :
i

19
 c2

c({T})

d*

Cs

c1

Figure 2.2: Demand level and satisfa tory ompromise set in two-dimensional
riteria spa e

Of ourse, we have (T )  d and therefore we have a satisfa tory

solution if   1. Moreover, it an be shown [205 that a minimum
solution of 
 = min max
i
f i (T )=di g
T
is a ompromise solution. Again, if  is less than or equal 1, the
ompromise solution is satisfa tory.

Hen e the problem of nding a satisfa tory ompromise solution is redu ed

to a s alar min-max optimization problem. This is also known as goal at-
tainment with a zero ideal point [90. Fig. 2.3 illustrates what best possible
solution is a hieved by min-max optimization in a two dimensional riteria
spa e.
The min-max optimization problem an be solved using standard nonlinear
programming methods applied to  as obje tive fun tion. However, the non-
smoothness of  due to the maximum fun tion may ause problems if gradient
based solvers are applied. In this ase, it is preferable to reformulate the
optimization problem in one of two ways:

1. The un onstrained min-max optimization problem with smooth riteria

i (T ) is equivalent to the onstrained problem [181
min max f i (T )=dg , minftg ; s:t: i (T )  t :
i
T i T;t
Solving the min-max problem in this way yields exa t solutions.

20
 c2

c({T})

d*

c*

c1

Figure 2.3: Satisfa tory ompromise found by min-max optimization

2. Approximate solutions are found if the fun tion  is approximated by a

smooth fun tion, as proposed in [138:

X
max
i
f i (T )=di g = lim 
!1 1= ln ( exp( i (T )=di ))
i

X
=  + lim
!1 1= ln ( (( i (T )=di
exp ))) :
i
This approximation formulation is well suited for numeri al omputation,
sin e the argument of the exponential is always less than or equal to zero.
The approximated min-max problem an be solved as an un onstrained
parameter optimization problem.

Of ourse, the fun tion  an be minimized dire tly if optimization methods

su h as dire t sear h methods are used, whi h do not require smooth obje tive
fun tions.

21
3. Eigenstru ture Assignment

Lester Faleiro , Jean-Franois Magni ,

1 2

Jess M. de la Cruz and Stefano S ala

3 4

3.1 Introdu tion

The theory presented here on erns the design hapters 17, 18, 19 and some
aspe ts of hapter 28. The main on epts of eigenstru ture assignment as a
design te hnique will be explored, in orporating a short explanation of how to
hoose a desired eigenstru ture based on design spe i ations. The mathemat-
i al methods used will also be summarised, and some omments given on the
use of eigenstru ture assignment.

3.2 Eigenstru ture Analysis

The equations that des ribe an air raft and their relation to the time response
of that air raft an be grouped together in matrix form:

x_ = Ax + Bu (3.1)
y = Cx + Du
where the most important of these matri es, A, des ribes the internal dynami s
of the air raft. The B matrix des ribes the distribution of the a tuator inputs
to the states of the air raft, and the C matrix denes how the states an be
observed as outputs of the system. D is usually zero for an air raft, though
non-zero matri es o ur when air raft a elerations are in luded in the outputs.
x is the state ve tor, u is the input ve tor and y is the output measurement
ve tor. It will be assumed that the system has n states, m inputs and p outputs.
A an be further de omposed into its onstituent eigenvalues and eigenve -
tors. The derivation of these an be found in any standard text on linear matrix
algebra. Let the n eigenvalues and eigenve tors of the system be dened by:
 = [1 : : : i : : : n and V = [v1 : : : vi : : : vn (3.2)
1
Department of Aeronauti al and Automotive Engineering and Transport Studies, Lough-
borough University, Loughborough, Lei estershire LE11 3TU, United Kingdom.
2
CERT ONERA, Dpartement d'tudes et Re her hes en Automatique, BP 4025, F31055
Toulouse Cedex, Fran e.
3
Dep. Informti a y Atomati . Fa ultad de Cien ias Fsi as. Universidad Computense.
28040 Madrid, Spain.
4
Flight Control and Me hani s department, Centro Italiano Ri er he Aerospaziali, 81043
Capua, Italy.

22
where
AV = V  (3.3)

The eigenve tor set V is a basis set for the state spa e x; thus any ve tor in
the state spa e an be expressed as a linear ombination of the eigenve tors of
the air raft system. These eigenve tors are also alled the right eigenve tors of
the system. The left, or dual basis eigenve tors of the same system are given
by W , where
W T = [w 1 : : : w i : : : w n ; W A = W (3.4)

Solving the state-spa e equations given in (3.1) yields an expression for the
time response that an be found in most standard ontrol texts:

n
X n
X Z t
y(t) = Cvi wTi ei t x0 + Cv i wTi ei (t  )Bu( )d (3.5)
i=1 i=1 0
It is lear from this equation that there are two omponents to the time
response. The rst is dependent on the initial onditions of the system, and
is alled the homogeneous omponent; the se ond is dependent on an input to
the system, and is alled the for ed omponent. The entire time response of a
linear system thus depends on four variables:

The eigenvalues of the system

The eigenve tors of the system
The initial onditions of the system
The inputs to the system

Ea h of these plays a part in the determination of the time response, and

di tates the overall ee t that modes and inputs play in the output response
of the system.
The homogeneous omponent of equation (3.5) an be written as

n
X
y(t) = Ci ei t vi (3.6)
i=1
where i are the s alars wTi x0 , i = 1 : : : n. This shows that the output response
is omposed of a linear ombination of eigenvalue-eigenve tor sets of the matrix
A. Ea h of these sets is alled a mode. In every mode the eigenvalue deter-
mines the de ay/growth rate of the response and the eigenve tor determines
the strength of the oupling of this mode with the outputs.
From (3.6) we an see that the oupling of theith mode with the j th output
is given by C j v i , where C j is the j th
row of C . If C j v i = 0, then equation
(3.6) shows that the i
th mode does not ontribute to the j th output; they have
been de oupled.
As an example of how the information about the nature of eigenstru ture
an be used, let us examine a simple linear representation of the longitudinal

23
dynami s of the RCAM model, in terms of four varying states of the sys-
tem. Mathemati ally, we an determine the time response of the system to
an arbitrary initial ondition, but this does not ne essarily give us a omplete
understanding of the system dynami s.

Mode Eigenvalue Damping ratio Frequen y (rad/s)

1 0:830  1:107i 0.6 1.38
2 0:011  0:126i 0.09 0.13

Table 3.1: Modes of the open-loop system

The eigenvalues of this nominal system are shown in Table 3.1. It an

be seen that although there are four states in the system, there are only two
modes in its dynami behaviour. It is known that the Phugoid and the SPPO
(Short Period Pi hing Os illation) are the two os illatory modes that hara -
terise air raft longitudinal motion, and that the Phugoid usually has a mu h
lower frequen y and damping than the SPPO. However, if these modes were in
any way un onventional, a knowledge of the eigenve tors alone would not be
su ient to understand the air raft.

States Mode 1 Mode 2

q 0:014 6 61:5 0:002 6 50:2
 0:010 6 8:3 0:0132 6 34:7
u 0:015 6 14:5 0:99 6 39:5
w 1 6 20:7 0:142 6 41:8
Table 3.2: Eigenve tors of the open-loop system (magnitude and argument)

The only way to ensure that ea h of the modes an be attributed to parti u-

lar air raft hara teristi s is by a subsequent examination of the right eigenve -
tors of the system. For this ase, these are shown in Table 3.2. The eigenve tors
for a mode are read verti ally down the table. It an be seen that Mode 1 is
hara terised by a large intera tion with w, the standard hara teristi of the
SPPO. Mode 2 is hara terised by a large intera tion with u, and omparatively
little with w, typi al of the Phugoid. The two modes an thus be designated
as 1. SPPO and 2. Phugoid.
In the time domain, the peak for ea h of these states will dier a ording
to the phase angles (arguments) of the elements of the eigenve tor, given in
degrees in Table 3.2. Note that usually the magnitude, rather than the phase,
in eigenstru ture assignment an be more easily visualised for the purposes
of design and analysis, so only the magnitudes will be used in eigenve tor
des ription from now on.
Additional information about the system an be obtained by using the left
eigenve tors to determine the ee t that ea h input has on ea h mode of the
system. These input oupling ve tors are given by the produ t of the left
eigenve tors and the input distribution matrix, W B. For the above example,

24
the input oupling is given below:

Mode t th
SPPO 85.4 19.3
Phugoid 31.5 13.7

This shows that the SPPO will be ex ited by a taileron input to a mu h

larger extent than a throttle input, and the Phugoid is the same. This quali-
tative eigenstru ture analysis is a tool that an be used to examine the nature
of the modes of a system qui kly. Classi al te hniques usually assume that a
knowledge of the system dynami s is readily available with the model. This
is a fair assumption, but may be ome redundant if more omplex modes are
involved in the open-loop system. Additionally, this te hnique of analysis is in-
valuable during the eigenstru ture assignment pro ess in examining the sour e
of design problems.

3.3 Eigenstru ture Assignment

It was shown in equation (3.6) that the output response of a air raft an be
des ribed by a representation involving its eigenvalues and eigenve tors. Thus,
if the eigenstru ture of the air raft an be manipulated somehow, we have a
means of altering its time response.
Various forms of dire t eigenstru ture assignment methodology exist, from
the rst tentative steps in output feedba k by Kimura [135 to their further
development by Andry, Shapiro and Chung [211 to urrent work su h as that
done by Sobel, Lallman and Shapiro [219, [221 and [220. In essen e, all these
methods are similar, and fun tion in mu h the same way. They all require the
designer to spe ify a set of eigenvalues and eigenve tors for the design, and
they all produ e a proportional gain matrix ontroller.

3.3.1 Determination of the desired eigenstru ture

The philosophy behind dire t eigenstru ture assignment is that whilst the de-
signer is able to spe ify a set of desired losed-loop eigenvalues d , she/he
is also able to spe ify exa tly whi h elements of the desired eigenve tors Vd
she/he would like to set to zero, where

d = [d1 : : : di : : : dp ; V d = [vd1 : : : vdi : : : vdp (3.7)

This an be illustrated by the set of eigenve tors shown in Table 3.3. We

would perhaps like the SPPO mode of response to be unae ted by forward
velo ity and pit h angle, and vi e-versa. We therefore spe ify these elements
in the desired losed-loop eigenstru ture to be zero. We are un on erned with
the values of the remaining elements, designated by an 'x'. A similar situa-
tion o urs with the Phugoid eigenve tor. This pro ess is ee tively assigning
elements of vi in (3.6) to zero.

25
 States SPPO Phugoid
q x 0
 0 x
u 0 x
w x 0

Table 3.3: Example of desired losed-loop eigenve tors

The ontrol design problem an thus be stated as follows: Given a set of

d and a orresponding set of desired eigenve tors V d , nd
desired eigenvalues
an m  p matrix K su h that the eigenvalues of the losed-loop system matrix
(A + BKC ), obtained when using the output feedba k ontrol equation
u = Ky; (3.8)

in lude d as a subset, and the orresponding eigenve tors of (A + BKC ) are

as lose as possible to the respe tive members of the set V d.

3.3.2 The a hievable ve tor spa e

Now, from the eigenve tor equation of the losed-loop system:

(A + BKC )vi = i vi ; i = 1 : : : p (3.9)

Avi i vi + BKCvi = 0 (3.10)

 
  vi = 0
A i I B (3.11)
zi
where
zi = KCvi (3.12)

So, for a non-trivial solution,

 
vi 
2 Ker A i I B

(3.13)
zi
 
and the rst n rows of the null spa e (Ker) of A i I B form the a hiev-
able ve tor spa e, N i . A se ond method that an be used to determine this
spa e an be derived from (3.10):

vi = (A i I ) 1 BKCvi (3.14)

Dene
N i = (A i I ) 1 B (3.15)

and now the losed-loop eigenve tors should omply with

26
 vi = N i zi (3.16)

in order to obtain the required eigenvalues. The a hievable eigenve tors vai must
lie in the subspa e spanned by the olumns of the matrix N i . Expanding this
example into more general terms, the subspa e des ribed by N i is of dimension
m. On e the desired eigenvalues have been hosen, the range spa e of matri es
Ni onstrains the sele tion of the losed-loop eigenve tors. In general, the
desired eigenve tors v di will not reside in the a hievable eigenve tor spa e. In
order to have the resulting eigenve tor as lose as possible to the desired one,
an optimum hoi e is made by proje ting the desired eigenve tor onto the
a hievable spa e, N i.
This is illustrated diagrammati ally in gure 3.1 for a simple three dimen-
sional system.

Dimension 3 desired vector achievable vector

v id v ia

This vector space, defined by

the null vectors, describes the
set of points over which the
desired eigenvalues can be realised.

Null space vectors

Figure 3.1: Representation of de oupling in a 3-dimensional state spa e

In this example, the desired ve tor an be hosen to de ouple a mode from

a dimension. As an example, say we want this mode to be de oupled from
Dimension 2. Thus, for this system, the only possible a hievable eigenve tor is
given by the interse tion between the null spa e (whi h is the only pla e where
the desired eigenvalue will be produ ed) and the Dimension 1/Dimension 3
plane (the lo us of points whi h does not ontain any omponent of Dimension
2). Sin e the desired eigenve tor vdi ontains desired de oupling information
(i.e. a zero in the Dimension 2 row), it will lie on the Dimension 1/Dimension
3 plane.
In real systems, this on ept an be expanded to de ouple modes from
air raft outputs. On e the desired eigenstru ture has been worked out, the
nal eigeve tors of the system an be produ ed.

3.3.3 Determination of the nal eigenve tors

A ording to [12, a reordering operator f
g
Ri an be dened su h that:
  
~i 
fvdi gRi = dli and fN i gRi N
= D (3.17)
i i

27
where li and di are the ve tors of spe ied and unspe ied omponents of vdi
respe tively. The rows of the null spa e Ni have been reordered in the same
way. The nal eigenve tor is given by (see [12)

y
vi = N i N~i li (3.18)

where (
)y denotes the pseudo-inverse.

It is also possible to determine the nal eigenve tors without the use of
proje tion. For ea h desired eigenve tor, the de oupled elements are integrated
into a row ve tor gi su h that if vdi = [x x 0 xT , gi = [0 0 1 0T . Thus,
gi vi = 0, sin e the nal eigenve tor should also have the relevant elements
de oupled. Thus, equation (3.11) an be rewritten as

  
A i I B vi
=0 (3.19)
gi 0 zi
and for a non-trivial solution,

   
vi 2 Ker A i I B (3.20)
zi gi 0
This ve tor in the null spa e an now be suitable partitioned and its rst n
entries an be used to form vai

3.3.4 Determination of the feedba k gain

These eigenve tors vi an now be grouped into the eigenve tor matrix V . The
ve tors zi ( omputed together with vi using (3.20)) are grouped into the matrix
Z. From (3.12) the feedba k gain satises

KCV = Z (3.21)

Usually, the number of olumns of V and Z is equal to the number of outputs

p, therefore
K = Z (CV ) 1 (3.22)

If the number of olumns is larger than p, a dynami fedba k an be used as

detailed in 3.4.4.
When vi is omputed as in (3.18), the orresponding ve tors zi an be found
easily in order to solve (3.22). However the resulting stati feedba k gain matrix
an be determined dire tly by substituting V into equation (3.9), whi h an be
rearranged to give:

K = B y (V  AV )(CV ) 1 (3.23)

Other ways of al ulating the gain matrix for numeri al e ien y and in the
ase of matrix non-invertibility have been des ribed in the literature ([12,
[133), and an be used instead of equation (3.23) if desired.

28
3.4 Robustness to Parameter Variation
Standard eigenstru ture assignment, as des ribed in previous se tions, takes
performan e and de oupling into a ount, but does not relate to any robustness
requirements. Four dierent, and sometimes o-operative, ways of ta kling this
problem have been pursued with the RCAM problem.

3.4.1 Open-loop ve tor proje tion

It has been shown by Wilkinson in [256 and [160 that for a perturbation in
the losed-loop matrix (A + BKC ) given by
(A + BKC ), the orresponding
rst order perturbation in the relevant eigenvalue is given by:

i = wi
(A + BKC )vi (3.24)

where w i and v i are normalized su h that w i v i = 1. On the assumption that

the open-loop eigenvalues do not vary a lot with parameter variation, (3.24)
shows that any variation an be related dire tly to the eigenve tors of the
system. Thus, if the open-loop eigenve tors are used as the desired eigenve tors,
eigenvalue sensitivity to perturbation should not be deteriorated by feedba k.
This thesis is used in the RCAM design in hapter 19.

3.4.2 Iterative assignment

Kautsy et al. [133 proposed using iterative eigenstru ture assignment to de-
rease the sensitivity of an eigenvalue in a state-feedba k ontrol system. An
iteration is used in whi h the ve tor vi is repla ed by a new ve tor with maxi-
mum angle to the remainder of the urrent right eigenve tor spa e V i for ea h
i = 1; 2; : : : ; n in turn. The new ve tor is obtained, letting:

V i = [v1 : : : vi 1 vi+1 : : : vn (3.25)

wi (ith left eigenve tor) is orthogonal to V i , and the new v i is found by pro-
je tingwi (now ee tively the desired ve tor for the ith mode) into N i (whi h
ontains the a hievable right eigenve tor spa e):

N i N Ti wi
vi =
k N Ti wi k2
(3.26)

thus giving a ve tor that is as orthogonal as possible to the urrent spa e

whilst retaining the desired eigenvalues of the losed-loop system. This means
that a perturbation in any of the elements of the remaining eigenve tors due to
parameter variation will not ae t the urrent mode. The iteration is ontinued
until the redu tion in the ondition number of the V matrix is less than some
toleran e. This is be ause the ondition number of the matrix ( ) is a measure 
of the overall sensitivity of the system. At the end of this iteration, a V matrix
for a minimum sensitivity solution remains. Ba k substitution of this matrix
into equation (3.23) produ es a feedba k gain matrix.

29
 Of ourse, altering the eigenve tors in this way does inevitable result in
a loss of performan e. The pro ess of de omposition and proje tion would
result in a loss of desired de oupling. However, using the null spa e des ribed
in equation (3.20) an help to over ome this problem, as the null spa e itself
ontains the de oupling required. A further des ription of the use of this pro ess
is given in [77.

3.4.3 Stability margin improvement

A se ond riterion in use is a measure of loop robustness in terms of gain and
phase margins. If the air raft is represented by G(s), a variety of loop transfer
fun tions an be used to determine losed-loop system robustness.
The singular values of the sensitivity fun tion S = (I + L) 1 , the om-
plementary sensitivity fun tion T = L(I + L) 1 and the balan ed sensitivity
fun tion S + T , where L is the open loop gain matrix, an be used to measure
the stability margins for multiloop feedba k ontrol systems ([152, [50 and
[28). The design pro edure in hapter 18 uses these measures, and the design
in [77 uses similar ones.
The fun tions S and T may be al ulated at the a tuator inputs or at the
sensor outputs. At the inputs, L = KG and at the outputs, L = GK . The peak
value of the maximun singular value ( ) of S , T or S + T gives a robustness
guarantee for all frequen ies. The formulae applied to omputing the stability
margins using the sensitivity fun tion are the following:

a = 1=(S )
Gain Margin = [1=(1 + a); 1=(1 a) (3.27)
Phase Margin = 2sin 1 (a=2)
where a  1. The gains of the loops may thus be perturbed simultaneously by
gains
satisfying 1=(1 + a) <
< 1=(1 a) without destabilising the losed
loop system. Similarly, the feedba k loops may be perturbed simultaneously
by phases  satisfying j  j< 2sin 1 (a=2) without destabilising the losed loop
system. The best possible gain and phase margins are obtained when  (S ) = 1,
in this ase the gain margin is [ 6 dB, +1 dB and the phase margin is 60 .
o
Similar margin equations an be devised for the T and S + T . These stability
margins are known to be onservative, and a better approa h is obtained by
repla ing the maximum singular value  with the stru tured singular value 
[44, [28.
The above des ription gives only a measure of robustness. In order to use
this information in a design synthesis, an iterative loop whi h ontains the
eigenstru ture assignment design pro ess, but updates the hoi e of eigenvalue
and eigenve tor an be used. This pro ess produ es variable results, depending
on the air raft and the initial design spe i ations, but has nonetheless been
found to be useful. Previous examples of the use of these stability margins to
improve robustness of air raft ontrol systems an be found in [178 and [76.

30
3.4.4 A multimodel approa h
A fourth way of improving the robustness of an eigenstru ture assignment
design is to use the multi-model approa h des ribed in [150. The RCAM
design des ribed in Chapter 19 uses this method.
It relies on produ ing a bank of linear air raft models at dierent operating
points. These models are denoted (Ai ; B i ; C i ) i = 1 : : : p. Extra freedom to
improve robustness is introdu ed with the multi-model approa h. Instead of
assigning all the available eigenstru ture to one linear model, a dierent model
may be used for ea h assignment. Thus, models with parti ularly sensitive
eigenvalues an be isolated, and the relevant eigenvalue-eigenve tor pair an
be re-assigned to improve the robustness of a parti ular mode on a parti ular
model.
Thus, for ea h eigenvalue in turn, hoose i and a model (Ai ; B i ; C i ) then
solve for v i , ti :
  
Ai i I B i vi
gi 0 zi = 0 (3.28)

First ase: the number of eigenve tors to be assigned is equal to the number
of outputs, solve for K by using:

K [C 1 v 1 C 2 v 2 : : : C p v p = [z 1 z 2 : : : z p (3.29)

K = [z1 z2 : : : zp [C 1 v1 C 2 v2 : : : C p vp 1 (3.30)

Se ond ase: more ve tors need to be assigned. It is ne essary to use a dynami

feedba k. Let K (s) denote the transfer fun tion matrix of the feedba k. In [150
is justied the fa t that we have to solve for K (s)
K (1 )C1 v1 = z1 ; K (2 )C2 v2 = z2 ; : : : (3.31)

Note that now, the assigned eigenvalue i appears in the equation. Finding a
solution to (3.31) is far more di ult than in the previous ase (see [150, [161
for details.)

3.5 Con lusions

This hapter has shown that the main pro ess of eigenstru ture assignment an
be broken up into two. The rst, and arguably most important, element is the
spe i ation of eigenstru ture based on the designers requirements and expe-
rien e. The se ond is the mathemati al pro ess of eigenstru ture assignment
itself.
This latter pro ess onsists of nding an a hievable eigenve tor spa e whi h
will produ e the desired losed-loop eigenvalues whi h have been spe ied for
performan e. Ve tors an then be hosen from this spa e to give required de-
oupling. Additional manipulation to redu e eigenvalue sensitivity an also be
employed. Robustness an best be a hieved by using eigenstru ture assignment

31
as a part of a large design strategy. Goal attainment, the use of singular values
and multi-model design have been des ribed as used for the RCAM problem.
Additionally, eigenstru ture an be further manipulated to give dynami
ontrollers, whi h have been des ribed for both the point design [77 and the
multi-model design [55. This is advisable in ases where additional design
freedom is required.
Despite all the versatility and potential visibility of the method, eigenstru -
ture assignment is most useful as a tool within a fuller design environment,
thus allowing the attainment of good performan e, de oupling and robustness
in the resulting ontrol system.

32
4. Linear Quadrati Optimal Control

Fran es o Amato 1, Massimiliano Mattei 1

and Stefano S ala 2

4.1 Introdu tion

Linear quadrati optimal ontrol is ertainly the most widely applied modern
ontrol te hnique. The fundamentals of this theory, whi h date ba k at least
to the Fifties (see the germinal paper [131 and the bibliography therein) an
be found in the Spe ial Issue on the LQG problem [1 whi h appeared as an
IEEE Transa tion on Automati Control in 1971; sin e then, many books have
been written on this subje t (see among others [10 and [140).

This ontrol te hnique allows the designer to take into a ount both re-
quirements on the amplitude of the ontrol inputs and the settling time of the
state variables; moreover, when onsidering innite horizon optimization and
provided that the weighting matri es are suitably hosen, an important feature
of LQ ontrol is that the resulting losed-loop system exhibits very good guar-
anteed multivariable stability margins. Many appli ations of the LQ theory
have been performed in the aeronauti al eld. One of the most important is
ertainly the design of the ight ontrol system of the AFTI/F-16 air raft by
General Dynami s (see [70).

When the omplete state is not available for measurement and some or all
of the measures are ae ted by noise, one an use the Kalman optimal ltering
theory [1 (whi h turns out to be the dual of the LQ optimal ontrol theory) to
design an observer of the state variables; however the robustness margins are no
longer guaranteed in the presen e of an observer. If sensor noise is absent or one
does not are about it, it is possible to use the degree of freedom on the design of
the observer to re over the LQ robustness margins; this is the elebrated Loop
Transfer Re overy (LTR) te hnique (see [226), whi h, however, an be applied
only when the plant under onsideration is minimum phase. Appli ations of
the LTR in the aereonati al eld an be found in [64, [203, and [249.

Finally in [231 some appli ations in aeronauti s of the linear quadrati

optimal stati output feedba k ontrol, developed in [172, are provided.

1
Dipartimento di Informati a e Sistemisti a, Universit degli Studi di Napoli Federi o II
via Claudio 21, 80125 Napoli, Italy, Tel.+39(81)7683172, Fax+39(81)7683686
2
Centro Italiano Ri er he Aerospaziali Via Maiorise, 81043 Capua (CE), Italy
Tel.+39(823)623949, Fax+39(823)623335

33
4.2 Plant Model Requirements and Controller
Stru ture
Let us start by onsidering the linear time-invariant plant

x_ = Ax + Bu u x(0) = x0 (4.1)

where, as usual, x(t) 2 IR n is the state and u(t) 2 IR m is the ontrol. The
steady-state Linear Quadrati (LQ) optimal ontrol problem an be stated as
follows:

Problem: given Q  0 and R > 0, nd, if existing, the ontrol law u : t 2

[0; +1) ! IR m whi h minimizes the ost fun tion:

Z +1  
J (u) = xT (t)Qx(t) + uT (t)Ru(t) dt : (4.2)
0
If the pair (A; Bu ) is stabilizable the problem is solvable and the optimal
ontrol law turns out to be a state feedba k ontrol law in the form

u(t) = Kx(t) (4.3)

therefore we often talk of Linear Quadrati State Feedba k (LQSF) optimal

ontrol law; the optimal gain matrix K is given by

K = R 1 BuT P (4.4)

where P is the unique positive semidenite solution of the algebrai Ri ati

equation
AT P + P A + Q P Bu R 1 BuT P = 0 : (4.5)

Finally the value of J (
) orresponding to the minimum is

Jopt = xT0 P x0 : (4.6)

Let us onsider the losed-loop system in Figure 4.1 given by the onne tion
of (4.1) and (4.3). As shown in [152 and [202, if the weighting matrix R is
hosen in diagonal form, this system exhibits, at the plant input, guaranteed
lower and upper multivariable gain margins of 1=2 and +1 respe tively; more-
over, the guaranteed lower and upper multivariable phase margins are 60o and
+60o respe tively. Therefore LQSF optimal ontrol systems, provided that the
weighting matrix R is properly hosen, have good robustness properties; this
fa t has further en ouraged ontrol engineers in appli ation of this te hnique
in several elds.
Now we assume that not all states are available for measurement and that
some or all of the measures are ae ted by white noise

x_ = Ax + Bu u + Bw w (4.7a)

y = Cx + m (4.7b)

34
 u -1 x
Bu (sI-A)
-

Figure 4.1: LQSF system

where y(t) 2 IR r and

ww ( ) = w (t  ) (4.8a)

mm ( ) = m (t  ) (4.8b)

are the auto ovarian e fun tions of the sto hasti pro esses w and m; we assume
that m is stri tly positive denite.
The steady-state Linear Quadrati Estimator (LQE) problem an be stated
as follows: Find a linear state estimator

x^ = L(u; y) (4.9)

whi h minimizes the steady-state mean square re onstru tion error

 T 
Ex (L) = t!lim
+1 E e x (t)e x (t) (4.10)

where
ex (t) = x(t) x^(t) : (4.11)

If the pair (A; 1w=2 ) is stabilizable and the pair (A; C ) is dete table, the
estimator problem is solvable; moreover the optimal estimator (whi h takes
the name of Kalman Filter) is a dynami system whi h possesses a Luenberger
observer stru ture

_ = A + Bu u + L(y C ) (4.12a)

x^ =  (4.12b)

where the optimal gain matrix L is given by

L = C T m1 (4.13)

and  is the unique positive semidenite solution of the algebrai Ri ati equa-
tion
A + AT + Bw w BwT C T m1 C  = 0 : (4.14)

35
Finally the value of the ost fun tion orresponding to the optimum is given
by
Exopt = tr() : (4.15)

It is readily seen that the LQ and the LQE problems are duals of ea h other.
An immediate onsequen e is that, if we onsider the losed-loop system in
Figure 4.2, this system exhibits at the output, the same robustness margins of
the LQSF system.

^
-1 x
(sI-A) C
-

Figure 4.2: LQE System

u x y
-1
Bu (sI-A) C
-

K Bu

^ + +
x -1 +
(sI-A) L
-

Figure 4.3: Controller-Observer Stru ture for Feedba k.

Now onsider the deterministi version of system (4.7)

x_ = Ax + Bu u (4.16a)

y = Cx (4.16b)

a well known result, the so- alled Separation Prin iple, states that, if one de-
signs a state feedba k gain K with A+Bu K Hurwitz, and a Luenberger observer

36
in the form (4.12) with A + LC Hurwitz, the losed-loop system depi ted in
Figure (4.3) and des ribed by the equations

x_ = Ax + Bu u (4.17a)

_ = A + Bu u + L(y C ) (4.17b)

y = Cx (4.17 )

u = K (4.17d)

is asymptoti ally stable; moreover, the eigenvalues of (4.17) are those of A+

Bu K and those of A + LC . Now assume that K and L has been designed
following an LQ optimal ontrol and Kalman Filter estimator philosophy re-
spe tively; we know from the above dis ussion that the LQ s heme without
observer in Figure 4.1 is robust at the plant input and that the LQE s heme
without state feedba k in Figure 4.2 is robust at the plant output. What an
we say about the robustness of the whole LQ-LQE s heme of Figure 4.3? The
answer, as shown by a ounter-example in [57, is, in general, nothing.
This last point introdu es the LTR robust ontrol te hnique, whi h is a
methodology to re over, in a ontroller-observer framework, the LQ (or the
LQE) robustness margins. Assume that the number of inputs is equal to the
number of outputs, that is m = r (if this hypothesis is not fullled and m < r we
an introdu e further  titious inputs), and that we are interested in obtaining
good performan e in terms of amplitude of the ontrol inputs and settling time
and, at the same time, good robustness properties at the plant input (a tuators)
in the s heme of Figure 4.3. We pro eed in the following way: rst the matrix
K is designed following equations (4.4) and (4.5) (after suitable matri es Q and
R have been hosen); then the observer gain L is hosen in su h a way that
the desired LQ margins are obtained at the plant input; the last part of this
se tion is devoted to detail the pro edure to design su h L. This pro edure is
known as LQG/LTR.

Assumption: the plant (4.16) is minimum phase.

Let L be the solution of an optimal estimator problem with  titious input

disturban e matrix and auto ovarian e matri es given by

Bw = I (4.18a)

w = Bu BuT (4.18b)

m =  2 I : (4.18 )

In this ase we have that

1
L( ) = ( )C T 2 (4.19)

where ( ) is the solution of
1
A( ) + ( )AT + Bu BuT ( )C T C ( ) = 0 :
2
(4.20)

It is shown in [141 that

1
lim L( ) = Bu U
 !0
(4.21)

37
where U is any orthonormal matrix. Using (4.21) and denoting G(s) = C (sI
A) 1 Bu as the transfer matrix of the plant and K (s;  ) as the transfer matrix
of the ontroller-observer, it is readily seen that

lim K (s;  )G(s) = K (sI

 !0
A) 1 Bu : (4.22)

Sin e K (s;  )G(s) is the open loop transfer matrix of the ontroller-observer
s heme in Figure 4.3, dened by opening the loop at the plant input, and K (sI
A) 1 Bu is the transfer matrix of the LQ ontroller in Figure 4.1, obtained
by opening the loop at the plant input, topologi al arguments lead to the
on lusion that the LQ robustness margins are asymptoti ally re overed at
the plant input in Figure 4.3, when  ! 0. From (4.21) follows that, when
 ! 0, the observer gain L( ) goes to innity; therefore, in pra ti al situations
one onsiders a given value of  , for example  = 1, and he k the degree of
satisfa tion of ondition (4.22) (this an be done by plotting and omparing
the singular values of K (s;  )G(s) and K (sI A) 1 Bu ). Then the value of  is
redu ed until the approximation of the limiting ondition (4.22) is satisfa tory
and ompatible with the fa t that faster and faster observers be ome more and
more transparent to sensor noise.
If we desire to re over the robustness margins at the plant output we an set
up the dual pro edure: rst design an optimal Kalman Filter and then design
an optimal LQSF ontroller using the  titious weighting matri es

Q = CT C (4.23a)

R = 2 I : (4.23b)

In this ase, the dual relations of (4.21) and (4.22) hold

1
lim K () =
!0
UC (4.24a)

lim G(s)K (s; ) = C (sI A) 1 L
!0
(4.24b)

whi h ensure the re overy of the Kalman Filter margins at the plant output.
We remember, however, that this last pro edure an only be applied when
r = m or r < m (in this ase it is ne essary to introdu e  titious outputs).
When the plant (4.16) is nonminimum phase, the full re overy of the sta-
bility margins annot be obtained; however, a partial re overy may result from
the modied LTR pro edure des ribed in [226.

4.3 Possible Design Obje tives and Design Cy le

Des ription
LQ optimal ontrol performs a trade-o between ontrol amplitudes and set-
tling time; this trade-o is strongly inuen ed by the hoi e of the weighting
matri es Q and R. Large values of R with respe t to Q will result in weak

38
 ontrol amplitudes and a slow regulation of the state variables; onversely we
have stronger ontrol amplitudes and a faster regulation.
For a system in the form (4.16) with m  r and robustness re overy at the
plant input (a tuators), the design y le is usually omposed of the following
steps:

Step 1 Choose the weighting matri es Q and R;

Step 2 Evaluate the time behaviour of states and ontrols;

Step 3 If the time behaviour is satisfa tory, go to Step 4, otherwise go to Step 1;

Step 4 Let  = ;
Step 5 Evaluate L( ) a ording to (4.19) and (4.20);
Step 6 Plot the singular values of K (s;  )G(s) and K (sI A) 1 Bu ; if the re overy
is not satisfa tory, set  = = , where > 1, and go to Step 5.

4.4 A Simple Design Example

We will now provide a numeri al example in whi h the LQ method has been
used to design a Proportional plus Integral feedba k multivariable a tion. This
stru ture resembles the one used in the HIRM ontrol s heme des ribed in
Chapter 28.
Let us onsider the linearized model of the longitudinal dynami s of the
HIRM air raft in straight and level ight (Ma h=0.40, altitude=10000 feet) in
the form

x_ = Ax + Bu u x(0) = x0 (4.25a)

y = Cx (4.25b)

where x = ( V  q  )T , u = ( T S engineF )T and y = x; we have the

following system matri es:

0
9:150 10 2 6:553 6:136 10 1 9:806 1
B 2:717 10 3 1:166 10 1 9:859 10 1 6:091 10 7 C ;
A = 
3:458 10 3 1:547 10 1 2:651 10 1 0:00 A
0:00 0:00 1:00 0:00
0
2:482 6:043 10 5 1
B 5:855 10 2 4:570 10 7 C ;
Bu = 
1:203 2:284 10 6 A C = I4 :
0:00 0:00
If we want to synthesize a ontroller whi h regulates velo ity and pit h rate,
we have to dene an auxiliary matrix:
 
Cr = 01 00 01 00

39
su h that
yr = ( V q )T = Cr x : (4.26)

We an now make referen e to the losed-loop s heme shown in Figure 4.4.

Considering that the state-spa e realization of the integrator is

x_ i = e (4.27a)

yi = xi ; (4.27b)

where e=r Cr x is the tra king error and r is the referen e signal, we have
the following losed-loop system state equation
   
x^_ = A + B
C
u Kp Bu Ki x^ + 0 r
0 I (4.28)
r
 
^ = xx
where x ; equation (4.28) an be rewritten as
i
 
x^_ = A^ + B^ K^ x^ + B^2 r (4.29)

where
     
A^ = A 0 ; B^ = Bu ; B^2 = 0I
Cr 0 0
are the state-spa e matri es of an auxiliary  ti ious system and

K^ = ( Kp Ki ) (4.31)

is the state feedba k gain whi h we are going to design with the LQ method.

r + e + u x y
1 Ki Linear Cr
s plant
- +

Kp

Figure 4.4: Closed Loop System

Now the problem is the hoi e of the weighting matri es Q and R for the
appli ation of the LQ te hnique to design the PI gain matri es. Indeed our
obje tive is to keep as low as possible, the tra king error e. This means that,
in the quadrati ost fun tion dened by the system ^ B^ ), the last two states,
(A;
whi h are related to the integrators, should be emphasized by in reasing the

40
relative entries of Q. In terms of the hoi e of R, a good trade-o between
performan e and ontrol energy must be found.
In Figure 4.5 the time response of the system is shown under a demand
of 5deg/se on q. The results obtained by dierent hoi es of the weighting
matrix R are ompared.
Q = diag(( 10 8 10 8 10 8 10 8 10 6 10 1 )) ;
8
< R0
R = 100 R0 R0 = diag ( 25 10 5 10 13 ) :
:
1000 R0
It is evident from the plots that, by in reasing the norm of the matrix R, the
ontrol energy, and onsequently the time response, des reases.

20 6

15 4
teta (deg)

q (deg/s)

10 2

5 0

0 2
0 5 10 0 5 10

30

20
R=R0
dts (deg)

10 R=100*R0
R=1000*R0
0

10

20
0 5 10

Figure 4.5: Simulation Results

41
5. Robust Quadrati Stabilization

Germain Gar ia1;2 , Ja ques Bernussou11 , Jamal Daafouz1;2

and Denis Arzelier

5.1 Introdu tion

A fundamental problem in ontrol theory is the robust stabilization problem
[56. From a pra ti al point of view, it is ne essary to hara terize a lass of
ontrollers whi h ensures, at least, asymptoti stability for the ontrolled un-
ertain system. A way to address this problem, is to extend the on ept of
Lyapunov stability to the ase of un ertain systems. The idea is to nd a sin-
gle Lyapunov fun tion for the ontrolled system from whi h a single ontroller
being dedu ed. When su h a Lyapunov fun tion exists, the system is said to
be quadrati ally stabilizable this is why the orresponding on ept is alled
quadrati stabilizability . Numerous papers deal with the quadrati stabiliza-
tion problem. For norm bounded un ertain systems whi h are entral in this
hapter, a solution is given in [193, [86 and onne tions between quadrati
stabilizability and H1 ontrol are presented in [134.
Stability is a minimum requirement and is not su ient in pra ti e when
a reasonable performan e level has to be obtained. A ommon and dire t way
to a ount for performan e is to put some onstraints on the losed-loop pole
lo ations leading to robust pole lo ation design. Another way is to dene a
performan e measure as H2 or H1 norms and, due to plant un ertainty, one
an at best minimize an upper bound on these norms. Su h approa hes are
referred to as guaranteed ost designs [87, [91. It is also possible to ombine
pole lo ation and guaranteed ost designs.
The rst problem addressed in this hapter is to nd a linear ontrol law
su h that the losed-loop system poles belong to the disk D(r; ) with enter
 + j 0 and radius r. The disk for pole lo ation an be hosen in su h a way
that a good ompromise between mode damping and speed is guaranteed. For
ontinuous systems, it su es to in lude it in a se tor lo ated in the left half
omplex plane. If  is a omplex mode for the ontrolled system, !n = jj, its
undamped natural frequen y,  = Re[, its damping fa tor and z = !n 1 ,
its damping ratio, then

8 2 D(r; ) !n <  + r;  < r ; z > r 1

1
LAAS-CNRS, 7 avenue du olonel Ro he, 31077 Toulouse Cedex 4
2
Also with INSA, Complexe S ientique de Rangueil, 31077 Toulouse Cedex 4

42
Another ontrol design problem whi h is dealt with in this hapter, onsiders
the disk pole lo ation ombined with a guaranteed H2 ost. When working in
the quadrati framework, two main approa hes are possible. The rst one (now
very popular) is hara terized by the use of an LMI formulation (Linear Matrix
Inequality) when writing the onditions for quadrati stabilizability, in luding
or not performan e requirements. Being linear with respe t to the unknown
matri es, the LMI formulation proposes a onvex parametrization of the robust
ontrollers. Among the good features asso iated with LMI, one an stress the
fa t that there exist e ient numeri al tools (industrial pa kages) working on
interior point methods. Another interesting feature lies in the ability of the
LMIs to aggregate several onstraints, provided these are written in terms of
LMIs (the ase for stru tural onstraints, integral quadrati onstraints, et . ).
The se ond approa h relies on the use of Ri ati type equations, a tool whi h it
is not surprising to nd here, in the framework of linear systems with quadrati
fun tions. E ient numeri al tools exist to solve parameter dependent Ri ati
equations. An advantage in expressing the onditions through Ri ati equations
is that ontrol interpretation is mu h easier. Usually in a Ri ati equation, two
weighting matri es, one for the states, and the other for the ontrol, appear.
This is the ase for the Ri ati equations arising in the quadrati stabilizability
problem. Their role and ee ts on the derived ontrol are well understood and
it is possible by a judi ious hoi e or by a trial and error method to sele t a
ontrol s heme satisfying some requirements. It should be noted that Ri ati
equations an be derived be ause the pole lo ation region is relatively simple
(a ir le). This is the reason why in the following, the quadrati approa h will
be illustrated by developing the results through the Ri ati framework. For
more omplex regions, no analyti al solutions in terms of matrix equations an
be obtained. But for a large lass of regions named LMI regions, the problem
an be solved by LMI te hniques. For more details, see [46.

5.2 Preliminaries
Throughout the hapter, the symbols 0; 1 respe tively denote the null matrix
and the identity matrix of appropriate dimension. M0 denotes the transpose
of the matrix M ( omplex onjugate transpose for omplex matri es). For
symmetri matri es A B , A < ()B means that the matrix A B is
and
negative denite (semidenite). (M ) denotes the spe tral radius of M and
 (M ) = (M 0 M )1=2 the maximum singular value. Let us onsider a ontinuous
system des ribed by :

x_ (t) = (A +
A)x(t) + Bu(t)

(5.1)
y(t) = Cx(t)
where A 2 IR
nn , B 2 IR nm , C 2 IR pn , u(t) 2 IR m is the input, x(t) 2
IR
n is the state. In order to simplify the following developments, the ase of
un ertainty ae ting only the dynami matrix A is onsidered, noti ing that
most of the given results an be extended to un ertain A and B matri es. These

43
results an be found in the given referen es. There are several ways to model
the un ertainty. One of the most popular is the following:

Norm bounded un ertainty

A = DF E (5.2)

where D2 IR
nr , E 2 IR
ln dene the stru ture of the un ertainty and the
modelling parameter un ertainty F belongs to the set :

F = fF 2 IR rl : F 0 F  1g (5.3)

In this way an ellipsoidal volume is dened as an un ertainty domain in the

hyperspa e of the entries of A, the nominal model being dened in the enter
of this hyperellipsoid. There exist some other ways to des ribe un ertainty.
We list below some examples whi h may be translated, after some elementary
transformations, into a norm bounded un ertainty. Their pra ti al interest is
dis ussed in some detail in the robust ontrol literature.

Bounded real un ertainty

The un ertainty term is written as :

A = DF (1 D0 F ) 1 E
where D 2 IR
nr , E 2 IR ln dene the stru ture of the un ertainty and the
modelling parameter un ertainty F belongs to F . D0 is a onstant matrix
satisfying 1 D00 D0 > 0. Then, we have
A +
A = A + D(1 D0 D0 ) 1 D0 E + D(1 D0 D0 ) 1=2
(1 D0 D0 ) 1=2 E
0 0 0 0
with
0
 1.

Positive real un ertainty

The un ertainty term is given by :

A = DF (1 + D0 F ) 1 E
where D 2 IR
nr , E 2 IR rn dene the stru ture of the un ertainty and the
modelling parameter un ertainty F belongs to the set :

Fp = fF 2 IR rr : F 0 + F  0g (5.4)

Moreover, D0 is a onstant matrix of appropriate dimension satisfying D0 +

D00 > 0. This ondition ensures that the matrix 1 + D0 F is invertible for all
F 2 Fp . Then, we have
A +
A = A D(D0 + D00 ) 1 E + D(D0 + D00 ) 1=2
(D0 + D00 ) 1=2 E
with

 1.
0

44
Stru tured un ertainty
The above dened un ertainties are alled unstru tured un ertainties in the
sense that they are dened through a single un ertainty matrix F whi h is
dened in a very global and general set. We an introdu e some stru tural
features on the un ertainty by dening multiblo k un ertainty terms, su h as

m
X

A =
Ai
i=1
where
Ai an be expressed by one of the following expressions

Ai = Di Fi Ei ; Fi 2 Fi
D00 i D0i > 0

Ai = Di Fi (1 D0 i Fi ) 1 Ei ; Fi 2 Fi ; 1

Ai = Di Fi (1 + D0 i Fi ) 1 Ei ; Fi 2 Fpi ; D0 i + D0 0i > 0
with Di and Ei are onstant matri es of appropriate dimensions and the sets
Fi and Fpi are dened respe tively like F and Fp . In this way, one an take
into a ount more pra ti al parametri un ertainty, but the onditions derived
in the sequel are only su ient.

In [85, the quadrati d stability on ept whi h is the ounterpart of quadrati

stability in the ontext of pole pla ement in a disk was introdu ed. We re all
below the denition this on ept.

Denition 5.1 The system x_ (t) = (A + DF E )x(t) is quadrati ally d stable

if and only if there exists a positive denite symmetri matrix P 2 IR nn su h
that :
(Ar + Dr F Er )0 P (Ar + Dr F Er ) P < 0 (5.5)

for all F 2 F with

p
Ar = (A 1)=r; Dr = D= r; Er = E= r
p (5.6)

This denition states that a system is quadrati ally d stable if there exists a
single matrix P satisfying (5.5) for all the systems in the un ertainty domain.
Pole lo ation is meaningful in the ase of non time-varying un ertainty, i.e. F
is a onstant matrix. It has been shown in [85 that equation (5.5) is in fa t a
su ient ondition for quadrati stability, the P matrix in (5.5) is a Lyapunov
matrix for the system (5.1), whatever F belongs to F. Furthermore, one may
expe t that for slowly varying un ertainty, the satisfa tion of (5.5) will ensure
a good transient behaviour for the ontrolled system. It is to be noti ed that
(5.6) is a dis rete Lyapunov inequality for the transformed system (5.6). In
fa t, a system is quadrati ally d stable if and only if the transformed system is
quadrati ally stable. This equivalen e allows to interpret the quadrati d sta-
bility as an H1 norm onstraint as is done for quadrati stability in [134. The
45
 ondition be omes: the system dened by x_ (t) = (A +
A)x(t) is quadrati ally
d stable if and only if

kEr (s1 Ar ) 1 Dr )k1 < 1 (5.7)

In the light of this result, the quadrati d stability problem and in the sequel the
quadrati d stabilization problem are equivalent to an H1 synthesis problem
that an be solved using for example an LMI formulation or a Ri ati equation
approa h. It is well known that the LMI te hniques are powerfull, parti ularly
in the ases where multiple onstraints and obje tives have to be taken into
a ount.
In the ases where analyti al solutions an be derived, for example a Ri ati
equation, the omplexity of LMI omputations remains higher than that of
solving a Ri ati equation [84.

5.3 Quadrati d Stabilizability by Output Feed-

ba k
In this se tion, we use the equivalen e between the quadrati d stabilization
problem and H1 ontrol synthesis to solve output feedba k pole lo ation in
a disk. The output feedba k quadrati d stabilizability is formalized in the
following denition.

Denition 5.2 The system is said to be quadrati ally d stabilizable via output
feedba k if there exists a linear time-invariant ompensator K (s) su h that the
losed-loop system ( losed by u = K (s)y) is quadrati ally d stable for all F 2 F .
From (5.7) written for the losed-loop system, the system is quadrati ally d
stabilizable via dynami linear output feedba k if and only if

kEr (s1 Ar Br K (s)Cr ) 1 Dr k1 < 1 (5.8)

As before, the problem an be solved using some standard te hniques. The

Ri ati equation approa h leads to

Theorem 5.3 The system (5.1) is quadrati ally d stabilizable by an output

ompensator if and only if, given R1 ; R2 ; Q positive denite symmetri ma-
tri es of appropriate dimensions, there exist  > 0 and two positive denite
symmetri matri es X 2 IR
nn ; Y 2 IR nn satisfying :

A0r (X 1 + Br (R1 ) 1 Br0 Dr Dr0 ) 1 Ar X + Er0 Er + Q = 0 (5.9)

Ar (Y 1 + C 0 (R2 ) 1 Cr E 0 Er Q) 1A0 Y + Dr D0 = 0

r r r r (5.10)

with :
1 Dr0 XDr > 0 (5.11)

Y 1 Er0 Er > A0r (X 1 Dr Dr0 ) 1 Ar + Q (5.12)

46
 Condition (5.12) implies that (XY ) < 1. Under the previous onditions, a
ompensator is given by :

p
_ = (A + BK + rDKd) + L(y C) (5.13)
u = K
where:

K = (R1 ) 1 Br0 (X 1 + Br (R1 ) 1 Br0 Dr Dr0 ) 1 Ar

Kd = Dr0 (X 1 + Br (R1 ) 1 Br0 Dr Dr0 ) 1 Ar
L = (1 Y X ) 1 Ar (Y 1 + Cr0 (R2 ) 1 Cr Er0 Er Q) 1 Cr0 (R2 ) 1

5.3.1 Output d stabilization algorithm

The following algorithm to he k quadrati d stabilizabilty an be dedu ed from
the monotoni behaviour of the solutions of the previous Ri ati equations.

Step 1 : Choose positive denite symmetri matri es R1 ; R2 ; Q, for example R1 =

1; R2 = 1; Q = 1 and >0
Step 2 : Solve the two Ri ati equations of theorem 5.3. If the solutions are posi-
tive denite and satisfy (5.11) and (5.12), Stop. The system is quadrat-
i ally d stabilizable by output feedba k. Compute the ontroller with
formula (5.13). Otherwise go to step 3.

Step 3 : Take  = =2. If  is less than some omputational a ura y 0 , Stop. The
system is not quadrati ally d stabilizable by output feedba k. Otherwise
go to step 2.

It is obvious that the above algorithm onverges for some 0 > 0 in a nite
number of steps. 0 has to be hosen su iently small. To solve the Ri ati
equations some standard algorithms an be used.

5.4 Quadrati d Stabilizabilty and Guaranteed

Cost
In this se tion, the results of robust pole lo ation in a disk are ombined with
another spe i ation requirement expressed through an H2 norm of a transfer
matrix from an external perturbation to a ontrolled output. In fa t, this
problem an be seen as a robust pole lo ation problem with the minimization
of an upper bound on a linear quadrati ost (multi obje tive ontrol design).
Let the un ertain system be des ribed by :

x_ (t) = (A + DF E )x(t) + Bu(t) + B1 w(t)

z (t) = C1 x(t) + D12 u(t) (5.14)
y(t) = C2 x(t) + D21 w(t)

47
where w is a disturban e, z 2 IR
s is a ontrolled output and F 2 F. All
matri es are onstant matri es of appropriate dimensions. We assume without
loss of generality that C10 D12 = 0 and B10 D21 = 0. Let us also dene :
Co = f > 0 : the onditions of theorem 5.3 are satisedg
and :
K = fK(s) given by (5.13) :  2 Co g

The ontroller K (s) an be written as :

_ (t) = H (t) + Ly(t)
(5.15)
u(t) = K(t)
where H ; L; K are given by theorem 5.3. The losed-loop system is obtained
by ombining (5.14) and (5.15).
2 Ae
z }| {

x_

=
6
6 A
+
0
 
0B 
H L 

0 1
_ 6
4 00
| {z } |
1{z0 }|
K 0
{z }|
C2 0
{z }
A B 3 H C
  7   
+
D F E 0
7 x + B1 w
0 | {z }5
7
 LD21
| {z } E | {z } | {z }
D   X B1
z = C1 D12 K

x
| {z } 
Ce
The transfer matrix from w(t) to z (t) is given by :
HF (s) = Ce [s1 Ae DF E 1 B1 (5.16)

If K (s) 2 K, the H2 norm of HF is expressed as :

kHF k22 = Tra e(Ce L (F )Ce0 ) = Tra e(B10 Lo (F )B1 ) (5.17)

where L (F ) and Lo (F ) are respe tively the ontrollability and observability

gramians solutions of :


(Ae + DF E )L (F ) + L (F )(Ae + DF E )0 + B1 B10 = 0
(Ae + DF E )0 Lo (F ) + Lo (F )(Ae + DF E ) + Ce0 Ce = 0
(5.18)

The problem solved in this se tion is the following :

Find K (s) and 0  < 1 su h that K (s) 2K and 8F 2

F , kHF k2  , being as small as possible.

48
 5.4.1 Guaranteed ost ontrollers
To solve this problem, note that the existen e of a ontroller satisfying the
onditions of theorem 5.3 is independent of the hoi e of the weighting matri es
R1 ; R2 and Q. In fa t, an appropriate hoi e for these matri es allows us to
solve the pole lo ation with guaranteed ost design problem. Let :

0 D12 + 1
R1 = D12 0 + 1
R2 = D21 D21 Q = C10 C1 + 1
where > 0 is a small parameter whi h prevents singularities. We have the
following results.

Theorem 5.4 Suppose that system (5.14) is quadrati ally d stabilizable by out-
put feedba k. Then:

i) Co 6= ;
ii) There exists 1 > 0 and P = P 0 > 0 su h that :
A0e P + PAe + 1 PDD0 P + 1 1E 0 E + Ce0 Ce + 1 = 0 (5.19)

iii) Let Eo1 = f1 > 0 : equation (5.19) has a solution P > 0g. For all  2
Co , 1 2 Eo1 , and F 2 F , we have :
P (1 )  Lo (F ); 8F 2 F
iv) For all  2 Co , K(s) is a guaranteed ost ontroller with
2 (1 ) =
tra e (B1 B10 P (1 )).

5.4.2 Optimization problem

The previous lemma suggests solving the following optimisation problem to
nd the best guaranteed ost ontroller in the sense dened above.

Min tra e [B1B10 P (1 )

 2 Co (5.20)
1 2 Eo1
We propose the following algorithm :

Algorithm. For a representative sample of values of  2 Co , do

Step 0 : Initialize
 (1020, for example)
Step 1 : Take  2 Co and ompute the orresponding ontroller using theorem
5.3.

Step 2 : For the ontroller obtained in step 1 , solve:

2 = A rg fMin tra e[B1 B10 P (1)g (5.21)

1 2 E o1

49
Step 3 : If <  , H H; L L; K  K;  , go to step 2.
Else go to step 1.

It an be shown that in step 2 , tra e[B1 B10 P (1 ) is a onvex fun tion with
respe t to 1 over Eo1 and then this optimization problem an be solved by a
one-line sear h algorithm.

5.5 Pra ti al Considerations

 To apply the method presented in this hapter, the rst step is to derive
an un ertain model for the system. Usually a nominal model is available (lin-
earization) and the un ertainties result from parameter variations, high fre-
quen y phenomena or non-linear ee ts. The term
A an be obtained by an
a priori knowledge of the range of parameter variations. The D matrix dis-
tributes un ertainties over the olumns of A and the E matrix over the rows
ofA. The size of the un ertainty is adjusted by an appropriate s aling on the
D and E matri es. For the RCAM design problem, the parameters whi h vary
are the mass and the entre of gravity.

 The high frequen y and non linear phenomena an be minimized by appropri-

ately shaping the sensitivity fun tions. A way to do this is to sele t judi iously
the weighting matri es R1 ; R2 and Q. Although no systemati method to
x these matri es exists, a trial-and-error approa h allows us to adjust them.
Theorem 1 hara terizes the lass of ontrollers whi h pla es the poles in a
disk and the weighting matri es an be used to nd in this lass, a ontroller
whi h satises other requirements. With no un ertainty, that is D and E equal
to zero, the ontroller derived from theorem 1 is lose to an LQG ontroller
dened on the triple (Ar ; Br ; Cr ). Then onsidering R1 =  1, R2 = 1 and
Q =  1, if    and  , a ontroller with small gains K and L will be
sele ted in the lass of disk pole lo ation ontrollers. These gains have a dire t
inuen e on the sensitivity fun tions. Then with ; and , it is possible to
shape the sensitivity fun tions. If D and E are not equal to zero, a similar
hoi e leads to the same ee ts. In pra ti e, a ompromise an be obtained by
a trial-and-error approa h.

 Another degree of freedom on erns the hoi e of the parameters dening

the ir le (; r). The values of these parameters are imposed by the settling
time and overshoot spe i ations, but there exists a ertain latitude on their
sele tion. If the radius of the ir le is too small, the problem is onstrained
and the lass of ontrollers too. In fa t, a trial-and-error approa h allows us to
obtain a satisfa tory ompromise.

 The last point on erns the onservative nature of the approa h. Consider
rst the unstru tured un ertainty. It is well known that the quadrati approa h
leads to onservative results be ause a xed Lyapunov fun tion is used for the
design. To alleviate this, it is possible to use parameter dependent Lyapunov
fun tion approa hes developed over the last few years. If un ertainty is stru -
tured, the onservatism is more important. A way to redu e it is to ombine a

50
 synthesis approa h with multipliers.

5.6 Con lusion

In this hapter, a robust ontrol design based on the quadrati approa h was
presented. The performan e requirements are onsidered following two dierent
paths. The rst one onsists of lo ating the losed-loop system poles in a disk,
the parameters dening the disk ( entre and radius) being hosen in a way that
ensured good transient behaviour.
The se ond one onsists of dening a ost fun tion (quadrati ) and minimising
a ost upper bound, leading to the well-known guaranteed ost design. In fa t,
these two means to express performan e are onsidered simultaneously in this
hapter.
The derived onditions expressed in terms of parameter dependent Ri ati
equations an be solved with available numeri al powerfull te hniques.

51
6. H1 Mixed Sensitivity
Mark R. Tu ker and Daniel J. Walker
1 2

6.1 Introdu tion

Classi al approa hes to feedba k design have for many years provided reliable
methodologies for designing ontrollers that are robust, but these te hniques
have not extended well to the multivariable ase.
Modern te hniques have subsequently looked at methods for designing mul-
tivariable robust ontrollers. H1 ontrol theory has been establishing itself
sin e the 1980's. The approa h is based on minimising over frequen y the peak
values of ertain system transfer fun tions that an be hosen by the design
engineer to represent design obje tives.
The H1 mixed sensitivity approa h allows the design engineer to meet
stability and performan e requirements in the presen e of modelling errors,
un ertainty and perturbations arising from disturban es or noise. Input and
output signals are shaped with frequen y dependent weights to meet robustness
and performan e spe i ations.
H1 mixed
This hapter is a tutorial hapter that will des ribe the theory of
sensitivity methods. H1 minimisation is des ribed, followed by a mixed sen-
sitivity one degree-of-freedom single input and single output design method.
Next a two degree-of-freedom multivariable mixed sensitivity design is onsid-
ered that in ludes disturban e inputs and a mat hing model.
H1 te hniques have been applied in the design hapters 21, 22, 23, 29, 30,
and 31 where mixed sensitivity as well as loop shaping and -synthesis methods
have been used. Loop shaping and -synthesis tutorials are given in hapters
7 and 8 respe tively.
More extensive treatment of H1 theory and appli ations an be found in
[215, 159, 266, 96, 61.

6.2 H1 Minimisation
Consider the standard problem of Figure 6.1, where z are the output errors or
r are the exogenous signals (referen e inputs and disturban es), e are the
osts,
measurements and u are the ontrols.

1
Engineering Department, University of Lei ester, University Road, Lei ester LE1 7RH,
United Kingdom. E-mail: mrtsun.engg.le.a .uk Tel: +44 116 252 2567/2874 Fax: +44 116
252 2619
2
Engineering Department, University of Lei ester, University Road, Lei ester LE1 7RH,
United Kingdom. E-mail: wjdlei ester.a .uk Tel: +44 116 252 2529 Fax: +44 116 252 2619

52
 r z
P

K
u e

Figure 6.1: Standard Problem

A ording to the signals, the open loop system P , of 6.1, is given as

      
z P11 P12 r r
= =P (6.1)
e P21 P22 u u
The losed loop relationship taken from r to z an be derived as
Tzr = P11 + P12 K (I P22 K ) 1 P21 (6.2)

The standard H1 optimisation problem is to nd a stabilising ontroller

K whi h is proper and minimises the supremum (lowest upper bound) over
frequen y of the maximum singular value of Tzr , the transfer fun tion from the
referen e inputs to the output errors or osts. That is, minimise

k Tzr k1 = Re(sup
s) > 0
 [Tzr (s) (6.3)

A stabilizing ontroller a hieving the minimum losed loop norm, k Tzr k1 =

opt , is said to be optimal. A stabilizing ontroller a hieving a losed loop norm
> opt is said to be sub-optimal.
P an be represented in state spa e form as
2 3 2 32 3
x_ A B1 B2 x
4 z 5 = 4 C1 D11 D12 54 r 5 (6.4)
e C2 D21 D22 u
It is worth noting that three spe ial ases of the standard plant P exist.
A 1-blo k problem is when D12 and D21 are square and su h a problem is
mathemati ally easier to solve than a 2-blo k problem where only D12 or D21
is square. A 4-blo k problem is when neither D12 or D21 is square and su h a
problem is the hardest to solve. Generally, all problems will require the solving
of two algebrai Ri ati equations, referred to as the ontrol and lter equations
respe tively.
In fa t the system of P needs to be onstru ted so that the following on-
ditions apply [92.

53
 1. (A; B2 ; C2 ) is stabilisable and dete table. This is required for the exis-
ten e of stabilising ontrollers.

2. D12 has full olumn rank and D21 has full row rank. This is su ient to
ensure that the ontroller is proper.

 
3.
A jwI B2 has full olumn rank for all w, enabling a stabilising
C1 D12  
A jwI B1
solution to the lter Ri ati equation. Also has full
C2 D21
row rank for all w enabling a stabilising solution to the ontrol Ri ati
equation.

The H1 optimisation an be solved using fun tions su h as hinfopt whi h

iteratively sear hes for the optimum opt and using hinf  whi h produ es a
solution for a parti ular suboptimal . These fun tions are available in the
Matlab Robust Control Toolbox [45.
The ontroller produ ed will be of the same order as the system P used.
A high order ontroller an easily result, and so ontroller redu tion is often
performed to eliminate unwanted or redundant states.
A more spe i system stru ture is now onsidered.

6.3 Mixed Sensitivity - One Degree of Freedom

d
+
r + e u + y
K G
-

Figure 6.2: Closed Loop Feedba k System

Figure 6.2 shows a simple losed loop feedba k system with referen e input
r, output y, output disturban e d, error signal e and ontrol signal u. From
this

y=d = e=r = (I + GK ) 1 = So (6.5)

This is dened as the output sensitivity. To a hieve small tra king error, good
transient behaviour and high bandwidth the output sensitivity needs to be
small at low frequen ies whi h an be a hieved by designing K to have high
gain at these frequen ies. Also

u=r = u=d = K (I + GK ) 1 = KSo (6.6)

= (I + KG) 1 K = Si K (6.7)

54
where Si = (I + KG) 1 is dened as the input sensitivity. (Note that in a single
input single output system So = Si ). To a hieve robustness it is ne essary to
a ommodate disturban es and un ertainties and it is also required to limit
high frequen y ontrol eort. For this KSo must be designed to be small at
high frequen ies whi h an be a hieved by designing K to have low gain at
these frequen ies. In order to meet the low and high frequen y onditions, the
design will in orporate frequen y dependent weights.

z1 z2
W1 W2
d
+
r + e u + y
K G
-

Figure 6.3: Closed Loop Feedba k System with Weights

Figure 6.3 shows the system of Figure 6.2 with added weights.
From this it an be written
2 3 2 3
z1 W1 W1 G  r 
4 z2 5 = 4 0 W2 5 (6.8)
u
e I G
whi h hen e denes the augmented plant P. The transfer fun tion Tzr an
be obtained using 6.8 in 6.2 and so the H1 problem is to nd a stabilising
ontroller that minimises

W1 So
k Tzr k1 W2 KSo 1 = (6.9)

If there is a bound on the H1 norm su h that k Tzr k1 < then

W1 So < (6.10)
W2 KSo 1
where > 0. This implies that

k W1 So k1 < k W2 KSo k1 < (6.11)

so it an be shown that

 [So < =[W1 1  [KSo < = [W2 1 (6.12)

Hen e the frequen y dependent weights W1 and W2 an be hosen to give the

bounds on the terms So KSo required to a hieve the required high and
and
low frequen y gains. In fa t W1 needs to be a low pass lter whilst W2 needs
to be a high pass lter.

55
 Broadly speaking, W1 and W2 determine the performan e and robustness
properties respe tively. For example, if the weights have been s aled so that
gamma is about one, it follows that W1 1 provides an upper bound on So . In
other words, W1 should be hosen to mirror the desired So , the latter being
determined largely by performan e requirements. Likewise, W2
1 will provide
an upper bound on KSo. This an be interpreted in terms of the losed loop
system's robustness to unstru tured additive model error; the larger KSo at
any given omplex frequen y s = jw, the smaller the additive model error that
will be required to destabilize the system. (This follows from the small gain
theorem). Conversely, knowledge of the likely size of the additive model error
di tates the safe upper bound on KSo. KSo an also be interpreted in terms
of the gain of the losed loop system from the output disturban es to a tuator
useage. It should also be noted that So an be interpreted as determining the
system's robustness to output inverse multipli ative perturbation. Thus the
larger So at a given frequen y, the less robust the system is to output inverse
multipli ative perturbation at that frequen y.
Given the state spa e representation of the plant and weights as

     
AG BG A1 B1 A 2 B2
G= W1 = W2 = (6.13)
CG DG C1 D1 C2 D2
then the state spa e form of P is onstru ted as

2 3
2 3 AG 0 0 0 BG
x_ 6 B1 CG A1 0 B1 B 1 DG 72 3
6 7
6
6 0 0 A2 0 B2
7
7 x
6 z1 7
= 6 76 7
6 7 6 74 r5 (6.14)
4 z2 5 6 D1 CG C1 0 D1 D1 DG 7
e
6
4 0 0 C2 0 D2
7
5 u
CG 0 0 I DG
Assuming that the onstru tion of this augmented plant meets the requirements
given in Se tion 6.2, then the H1 minimisation an be performed to produ e
a robustly stabilising ontroller. Note that a 2-blo k problem is being solved
as D21 is square.

6.4 Design Pro edure

The following is a simple pres riptive pro edure for designing a one degree-of-
freedom ontroller using mixed sensitivity.

1. Sele t the linearised plant model.

2. Sele t W1 1 and W2 1 to bound So and KSo .

3. Augment the plant with the weights W1 and W2 to form the augmented
plant P.

56
 4. Synthesise a sub-optimal ontroller or an optimal ontroller where the
H1 norm is minimised. The smaller indi ating a more robust design.
5. System analysis.

The results an be tuned by adjustments to the weights and so iterations of

the design loop an be performed as required.
(Note: The pole-zero an ellation phenomenon an o ur in this one degree-
of-freedom mixed sensitivity te hnique. Steps to prevent this possibly undesir-
able situation o urring an be found in [241. The subsequent two degree-of-
freedom approa h does not suer from this phenomenon).

6.5 Design Example

The following plant model is onsidered to represent the transfer fun tion from
the input voltage to the angular position of a simple motor.

1
G= (6.15)
s(s + 1)
W1 is sele ted to be an integrator, this will result in integral a tion in
the ontroller enabling good tra king and small steady state errors. A true
integrator annot be used as this would not onform to the requirements of
Se tion 6.2, so an approximate integrator is used. The gain of this weight
determines the losed loop bandwidth. The sele ted weight is

1
W1 =
s + 10 6
(6.16)

Weight W2 is sele ted to be a high pass lter. W2 must be proper to meet

the rank onsiderations of the augmented plant as required in Se tion 6.2. The
gain and bandwidth of the weight are hosen to allow low frequen y ontrol
eort but limit high frequen y ontrol eort. The sele ted weight is

20s + 4
W2 = (6.17)
s + 80
The augmented plant is onstru ted as in 6.14 and a sub-optimal ontroller
is synthesised realising a H1 norm bound of = 1:1 opt = 1:2642. An optimal
ontroller an give better results over the whole frequen y range, but this may
be a hieved through high frequen y or dire t terms in the ontroller. A sub-
optimal ontroller is generated without these possible unwanted terms at a
slight ost to the robustness and performan e.
The output sensitivity So frequen y response of Figure 6.4 shows the desired
low gain over the operating bandwidth so reje ting low frequen y disturban es.
At high frequen ies the gain is unity and around the bandwidth there is a
peak in the response. The smaller this peak, the more robust the design. The
magnitude of this peak determines the smallest unstru tured output inverse

57
multipli ative disturban e that will de-stabilise the system (see Appendix of
[242). The fun tion KSo is similarly analysed.
Figure 6.5 shows the frequen y response the step response for the losed
loop system.
Iterations of the design y le an now be performed to meet robustness and
performan e spe i ations as required.

5 10

0 0

5 10

10 20
Gain [dB]

Gain [dB]
15 30

20 40

25 50

30 60

35 70

40 2 1 0 1 2 3
80 2 1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10
Frequency [rad/s] Frequency [rad/s]

Figure 6.4: (a) Output Sensitivity Frequen y Response (b) Frequen y Response
of KSo

1.2

0.8
Units

0.6

0.4

0.2

0
0 1 2 3 4 5 6 7 8 9 10
Time [s]

Figure 6.5: Step Response of the Closed Loop System

6.6 Mixed Sensitivity - Two Degree of Freedom

So far it has been shown how a one degree-of-freedom ontroller an be pro-
du ed using mixed sensitivity te hniques. The method is now extended to a
two degree-of-freedom ontroller design. The system of Figure 6.6 shows a pos-
sible onguration, with referen e input ( r1 ), an output disturban e input (r2 )

58
 z2
W2
r2
+
r1 e1 u + + z1
G W1
e2 K -

Figure 6.6: Two Degree of Freedom Closed Loop Feedba k System

and two output osts ( z1 and z2).M is an ideal model to mat h

The system

 K
the losed loop system to. Controller is to be designed.

K an be partitioned as K = K1 K2 , su h that
 

u = K1 K2
 e1 (6.18)
e2
where K1 is a pre-lter and K2 is a feedba k ontroller. There are two methods
for generating the ontroller K . Firstly, K2 ould be synthesised to robustly
stabilise the loop against disturban es and un ertainty, and then K1 synthesised
to shape the losed loop to meet the performan e requirements. Su h a two
stage approa h an oer greater exibility and may produ e better results, but
the method is ompli ated to implement needing a two step design pro edure
and the resulting ontrollers are independent of ea h other and so overall are
of a high order. A simpler one stage method is to generate the ontroller K by
synthesising the feedba k ontroller K2 and pre-lter K1 together. Only one
synthesis is required, and the resulting ontroller is of a lower order as K1 and
K2 share the same state spa e. Here the one stage approa h is onsidered.
For this system the standard problem of Figure 6.1 an be formed. The
system is represented by
2 3 2 3
z1 W1 M W1 W1 G 2 3
6 z2 7 r1
6 7 = 6
6 0 0 W2 7 7 4 r2 5 (6.19)
4 e1 5 4 I 0 0 5
e2 0 I G u

whi h hen e denes P . It an also be shown that

 
Tzr = W1 (So GK1 M ) W1 So (6.20)
W2 Si K1 W2 K2 So
On e again the aim is to produ e a stabilising ontroller K that minimises
the H1 norm of Tzr . The four weighted fun tions to be minimised are the
dieren e between ideal and a tual systems, the output sensitivity, the ontrol
eort to the referen e inputs and the ontrol eort to the outputs.

59
 W1 and W2 are frequen y dependent weights sele ted as before. So using
this design enables robustness and performan e riteria to be met in orporating
performan e spe i ations in the mat hing model M. The state spa e form of
P an be formed using the denitions of G, W1 and W2 as given in 6.13 and
using the state spa e representation of the mat hing model

 
AM BM
M= (6.21)
CM DM
The state spa e form of P is

2 3
Ag 0 0 0 0 0 Bg
2 3 6 B 1 CG A1 0 B1 CM B 1 DM B1 B1 DG 7
x_ 6 72 3
6 7
6
6 0 0 A2 0 0 0 B2 7
7 x
6 z1 7 6 0 0 0 AM BM 0 0 76 7
6
6 z2
7
7
6
=6 76
76
r1 7
7
6 7 6 D1 CG C1 0 D1 CM D1 DM D1 D1 DG 74 r2 5
4 e1 5 6 7
6 0 0 C2 0 0 0 D2 7 u
e2 6 7
4 0 0 0 0 I 0 0 5
CG 0 0 0 0 I DG
(6.22)

Assuming that the onstru tion of this augmented plant with plant model
and weights meets the requirements given in Se tion 6.2, then the H1 minimi-
sation an be done to produ e a robust stabilising ontroller. It is noted that
a 2-blo k problem is being solved as D21 is square.

6.7 Design Pro edure

The pro edure for designing a two degree-of-freedom ontroller using mixed
sensitivity follows a similar pro edure to the one degree-of-freedom ase (6.4).
The two degree-of-freedom ase requires the additional sele tion of a desired
losed loop system response model.

6.8 Design Example

The following plant model has been sele ted.

2 3
1 (s + 5) 0
G= 4 0:1(s + 5) 10(s + 1) 5 (6.23)
s(s + 1)(s + 5) s(s + 5) 0
The rst and se ond outputs are to be mat hed to the desired model

 1 0

M = s2 +2s+1 1 (6.24)
0 s +2s+1
2

60
The ross oupling terms here are zero, so dening the requirement for the
losed loop system to be de oupled.
Next, the weights W1 and W2 are sele ted. These will be multivariable and
frequen y dependent. Weight W1 is sele ted as an integrator to provide good
tra king and small steady state errors. The gain on this term will determine
the error bound on the dieren e between the a tual and the ideal system, and
will also determine the bandwidth for output disturban e reje tion.

2
W1_1 = W1_2 =
s + 10 6
(6.25)

The third output is the rate of the rst output and will be fed ba k to
enhan e the ontrol and robustness. Low frequen y a tivity of the signal is
required to enable the tra king of the rst output. The weight sele ted is a
bandpass lter, sele ted to reje t disturban es around the bandwidth frequen y.

s
W1_3 = 2 (6.26)
s + 2s + 1
Weight W2 is sele ted to be a high pass lter and must be proper to meet
the rank onsiderations of the augmented plant as required in Se tion 6.2. The
gain and bandwidth of the weight are hosen to allow low frequen y ontrol
eort and to minimise high frequen y ontrol eort.

20s+4 0
 
W2 = s+80 20s+4 (6.27)
0 s+80
A suboptimal ontroller is now synthesised realising an H1 norm of =
1:1 opt = 0:5818.
Figure 6.7 shows the frequen y response of the singular values of error
between the a tual and the ideal responses. At low and high frequen ies, the
error is small. Small error at low frequen ies will give good mat hing to the
model resulting in small steady state errors. Around the bandwidth, the error is
largest, although it is less than unity. Redu ing the error around the bandwidth
will improve the overshoot and rise times of the losed loop system. Above the
operating bandwidth the desired response is for low gain, whi h is a hieved as
the error is small.
The output sensitivity So response is also shown in Figure 6.7, where there is
small gain at low frequen y for the ontrolled hannels, whilst the rate feedba k
has unity gain over these frequen ies. At high frequen ies the gain is unity and
around the bandwidth there is peak in the response. The smaller this peak,
the more robust the design (as dis ussed in 6.5. Fun tion K2 So and Si K1 are
similarly analysed.
For the losed loop system Figure 6.8 shows the response of the rst hannel,
its rate and the ideal response of the mat hing model and the response of the
se ond hannel and the ideal response of the mat hing model.
Iterations of the design y le an now be performed to meet robustness and
performan e spe i ations as required.

61
 0 5

0
20
5

40 10

Gain [dB]
Gain [dB]

15
60
20

80 25

30
100
35

120 2 40 2 1 0 1 2 3
1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10
Frequency [rad/s] Frequency [rad/s]

Figure 6.7: (a) Frequen y Responses of the Dieren e Between M and S0 GK1
(b) Output Sensitivity Frequen y Response

1.2 1

0.9
1 Ideal Output 1
0.8

0.8 0.7

0.6 Ideal
0.6 Output 2
Units
Units

Output 1 0.5

0.4
0.4

0.3
0.2 Output 3

0.2 Output 2
0
0.1

0.2 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
Time [s] Time [s]

Figure 6.8: (a) Step Response to the First Input of the Closed Loop System
(b) Step Response to the Se ond Input of the Closed

6.9 Con lusions

This hapter has given a brief tutorial on H1 ontrol theory and how it an
be applied in H1 mixed sensitivity one and two degree-of-freedom design pro-
edures. The H1 mixed sensitivity method lends itself well to systems whi h
are required to meet stability and performan e requirements in the presen e
of modelling errors, un ertainty and perturbations arising from disturban es
or noise. Un ertainty and disturban e an be expli itly in orporated into the
design and stability is guaranteed subje t to bounded perturbations, although
robust performan e is not. The frequen y domain pro edures are demonstrated
using simple examples. Weight sele tion an be made to a ount for model
un ertainty. If model un ertainty is unspe ied, then the weight sele tion is
broadly dened by robustness and performan e requirements. Additionally
in the two degree-of-freedom design, a model is in orporated that is dire tly
translated from the performan e requirements that the losed loop system is

62
required to meet. Generally, ontrollers are produ ed by iterative design pro e-
dures. The weights are sele ted and the robustness and performan e analysed.
Large order ontrollers an sometimes be generated, but in pra ti e it is usually
possible to a hieve signi ant order redu tion.

63
7. H1 Loop-Shaping
George Papageorgiou and Keith Glover , 1 1

Alex Smerlas and Ian Postlethwaite

2 2

7.1 Preliminaries
The gain of a matrix A 2 IR nm an be dened as kkAx k2
xk2 where x 2 IR is an
m
input ve tor and k:k2 denotes the Eu lidean 2-norm. It an easily be dedu ed,
after a few al ulations, that the gain of A will depend on the dire tion of the
input ve tor x. To see this we dene the singular value de omposition (SVD) of
a matrix (see pp. 32-35 in [266). For example the SVD of a matrix A 2 IR
22
is   
  (A) 0 v1 
A = u1 u2 ;
0  (A) v2 
2 IR 2
   
where u1; u2 ; v1 ; v2 and u1 u2 and v1 v2 are unitary matri es
 (A) denotes the maximum singular value of A and  (A)
(see p. 19 in [266).
the minimum singular value of A. Therefore, as

    
v1  v v  = 1 0
v2  1 2 0 1
if x = v1 then the gain of A will be  (A) with Ax in the dire tion of u1 .
Similarly, for x = v2 the gain will be (A) with Ax in the dire tion of u2 .
Dene the ondition number of A as
(A)
(A) = :
(A)
Hen e an ill- onditioned matrix, i.e. a matrix with a high ondition number,
has signi antly dierent gains in dierent dire tions. A round matrix is a
matrix with a ondition number near unity.
Assume that we are given a linear time invariant (LTI) stable system G
with m inputs and p outputs, i.e. G(j!) 2 Cpm .
I All inputs u(t) into the
system are assumed to have nite energy, i.e. a nite 2-norm. We dene the
2-norm of a signal u(t) as
Z +1
kuk2 = ( uudt) 2 :
1

1
1
Cambridge University Engineering Department, Cambridge CB2 1PZ, England
2
Lei ester University, Lei ester LE1 7RH, England

64
It is easy to prove (see pp. 18-25 in [61) that if kuk2  1 the energy of the
output y(t) will be bounded by
supfkyk2 : kuk2  1g = kGk1
where the 1-norm of G is dened as
kGk1 = sup (j!):
!
Therefore kGk1 denotes the maximum gain of G over all frequen ies and all
input dire tions. Consequently, if for example we wanted good disturban e
reje tion at the output of a plant we would try to minimise the 1-norm of
the output sensitivity. The 1-norm an be a onservative measure if we are
not interested in all input dire tions. This motivates the use of the stru tured
singular value  (see pp. 271-300 in [266).
If P is a given plant model, then P = M~ 1 N~ is a normalised left oprime
~
fa torisation of P where M ~
and N are stable rational transfer matri es satis-
fying the normalisation onstraint

N~ N~  + M~ M~  = I:
The motivation for using H1 te hniques to design robust ontrollers is
provided by the small gain theorem (see pp. 217-221 in [266). From the small
gain theorem it an be dedu ed that by minimising the 1-norm of a stable
1-norm sense, of the unstru tured
transfer matrix we maximise the size, in an
perturbation to whi h the system remains stable. Hen e a typi al H1 ontrol
problem would be to minimise the 1-norm of a transfer matrix, alled the
generalised plant, over all stabilising ontrollers. This optimisation problem
has an exa t solution [63. The transfer matrix we hoose to look at depends on
the type of un ertainty present in the plant and the performan e spe i ations.
The rst step of a typi al H1 design pro edure would be to de ide on the type
of un ertainty to be used (see table on p. 227 in [266). This is di ult and
requires good knowledge of the plant. Normalised oprime fa tor un ertainty
is the most general type of unstru tured un ertainty (see pp. 418-419 in [96).
The se ond step would be to hoose frequen y dependent weights a ording to
performan e spe i ations and solve the optimisation problem (see pp. 213-
245 in [266). Some well studied H1 ontroller design te hniques are H1
loop-shaping, the S/KS design pro edure (see Chapter 6) and -synthesis (see
Chapter 8).

7.2 Overview of the Design Pro edure

The H1 loop-shaping design pro edure [164 is des ribed below:
1. Shape G open loop with frequen y dependent weights W1 and W2 a -
ording to losed loop obje tives. The weighted plant, Gs = W2 GW1 , is
depi ted in Figure 7.1.

65
 w1 w2 z1
? W2  G  W1  ? 6

z2
- K1

Figure 7.1: The H1 loop-shaping standard blo k diagram

2. Minimise the 1-norm of the transfer matrix from disturban es w1 and

w2 to outputs z1 and z2 over all stabilising ontrollers.

     

w1 ! zz1

= K1 (I

Gs K1 ) 1 I Gs

= :
w2 2
1 I
1
(7.1)
Che k the a hieved . This gives a measure of how robust the desired
loop shape is.

3. Choose the position of K1 in the loop. Model redu e the ontroller and
design the ommand pre-lters.

4. Che k the time simulations and frequen y responses of the resulting

losed loop system to verify robust performan e. Reiterations may be
required.

7.3 Justi ation of the Set-Up

H1 loop-shaping developed by M Farlane and Glover [164 is a very intuitive
method for designing robust ontrollers as the notions of lassi al loop-shaping
readily arry through. The designer an spe ify losed loop requirements su h
as disturban e and noise reje tion by simply shaping the loop gains. An impor-
tant feature of H1 loop-shaping is that it enables the designer to push for the
best a hievable losed loop performan e subje t to a required level of robust-
ness. This is be ause the designer has ontrol over the ross-over frequen ies
of the loop gain singular values.
When setting up a robust ontrol problem a de ision has to be made about
the type of un ertainty to be used. This an be di ult as it requires good
knowledge of the plant. The value of in Equation (7.1) provides a level of
robust stability to oprime fa tor un ertainty whi h requires no assumptions
to be made about the open loop stability of the perturbed plant. Coprime
fa tor un ertainty is a general type of un ertainty similar to the single-input
single-output (SISO) gain and phase margins. When there is little detailed
knowledge about the un ertainty present in a plant H1 loop-shaping is a good
method for designing robust ontrollers.

66
 Performan e is spe ied by shaping the singular values of the plant G with
weights W1 and W2 , Figure 7.1. In [164 it is proved that if is not too large
(say < 4) the ontroller K1 does not alter the desired loop shape Gs very
mu h. Hen e shaping G orresponds to shaping the loop gains Gs K1 and
K1 Gs . All losed loop transfer fun tions an be bounded in terms of Gs , W1 ,
W2 and (see pp. 493-494 in [266). In [210 it is proved that there are no left
half plane pole/zero an ellations between K1 and Gs . This is be ause K1
an be written as an exa t plant observer plus state feedba k. Hen e H1 loop-
shaping ontrollers an be gain s heduled. This was done in [120. Left half
plane pole/zero an ellations are undesirable as they an limit the a hievable
robust performan e.
The ost fun tion minimised provides robust stability. A measure of the
robustness of the desired loop shape is given by .
Cal ulating the best
requires no -iteration [164. Appli ation of the method to real plants has
shown that a value of  = 1= between 0:2 0:3 is satisfa tory in the same way
that a gain margin of 6dB and a phase margin of 45 are for a SISO design.
o
A good value of  should guarantee that the required gain and phase margins
are a hieved. In [247 the phase and gain margins of a stable SISO losed loop
system, PM and GM respe tively, are related to the obtained .
GM  (1 + )=(1 ); P M  2 ar sin():
A lot of resear h has gone into H1 ontrol design te hniques. As a result a
great deal of powerful analysis and model redu tion tools are available to help
with the erti ation of su h ontrollers, e.g. the gap metri , Hankel norm
approximation, the  -gap and -analysis.

7.4 Classi al Loop-Shaping

Before being able to design ontrollers using the H1 loop-shaping design pro-
edure (LSDP) the designer must be ome familiar with the notions of lassi al
loop-shaping (see pp. 130-137 in [266). Loop-shaping allows the designer to
spe ify losed loop obje tives by shaping the loop gains. Closed loop obje tives
su h as disturban e reje tion at both input and output of the plant, noise reje -
tion, output de oupling and tra king an easily be spe ied when loop-shaping.

di d
r - - K
u- ?up- G - ? -y
6
?
6n
Figure 7.2: A typi al losed loop system

67
 Simple algebrai manipulations on the losed loop in Figure 7.2 give:

y =To (r n) + So Gdi + So d (7.2)

r y =So (r d) + Ton So Gdi (7.3)

u =KSo (r n) KSod Ti di (7.4)

up =KSo (r n) KSod + Si di (7.5)

where Lo = GK , Li = KG, S = (I + L)
1 and T = I S . It is simple to show
that So P = P Si and Si K = KSo .
Disturban e reje tion at the plant output y , Equation (7.2), an be a hieved
by making  (So ) =
1
(I +GK ) small to reje t d and  (So G) =  (GSi ) small to
reje t di . Similarly, disturban e reje tion at the plant input up , Equation
(7.5), an be a hieved by making  (Si ) =
1
 (I +KG) small to reje t di and
(Si K ) = (KSo ) small to reje t d. Sin e (pp. 31-33 in [96)
(KG) 1   (I + KG)  (KG) + 1
(GK ) 1   (I + GK )  (GK ) + 1;
we an dedu e that if (KG) > 1 and (GK ) > 1
1 1
 (KG)+1   (Si )   (KG) 1
1  (So )  1 :
 (GK )+1  (GK ) 1
Therefore for disturban e reje tion of d at y and di at up
 (So )  1 , (GK )  1
 (Si )  1 , (KG)  1:
Also, if  (GK )  1 or (KG)  1 and assuming that G and K are invertible
(this assumption is made for the purpose of illustration) then

  (So G)  (1K ) . Therefore for disturban e reje tion of di at y the singular

values of the ontroller should be high at low frequen ies.

  (KSo)  (1G) . This gives the input required to an el the inuen e of d

on up . This will be small if  (G)  1 but an not be set by the designer
and onstitutes a physi al limitation of the plant.

As designers we are not only interested in disturban e reje tion. For noise
reje tion (To ) must be made small at high frequen y. Typi ally noise is only
important at high frequen y. Note that noise reje tion at low frequen y on-
i ts with disturban e reje tion as T + S = I. Large loop gains outside the
bandwidth of G an make ontrol a tivity quite una eptable. This an be
dedu ed by examining Equation (7.4). Output de oupling and tra king are
equivalent to reje ting d at the plant output be ause T + S = I . Hen e for ing
S to zero for es T to the identity. Figure 7.3 illustrates the desired loop shapes.

68
  (L)

!l !h log !
 (L)

Figure 7.3: Loop gain boundaries

7.5 Choi e of Weights

In [164 it is proved that K1 does not modify the desired loop shape signi-
antly, i.e. the loop-shaping ontroller is well- onditioned. Hen e shaping Gs
a tually shapes both Gs K1 and K1 Gs . This is a very important result that
onstitutes the theoreti al justi ation of H1 loop-shaping. The fa t that K1
is well- onditioned is intuitive by examining the ost fun tion minimised and
noting that this transfer matrix an be written in two ways (see p. 485 in [266).
All transfer fun tions an be bounded in terms of Gs , W1 , W2 and as
mentioned in Se tion 7.3. For example the input and output sensitivity are
bounded by

(I + GK ) 1  minf (M~ s )(W2 ); 1 + (Ns )(W2 )g

(I + KG) 1  minf1 + (N~s )(W1 );  (Ms )(W1 )g;
where K = W1 K1W2 . Therefore, as  (M~ s ) = (Ms ) = ( 1+(W12 GW1 ) ) 21 at
ea h frequen y, large low frequen y gain of Gs results in reje tion of d at y
and di at up . Choosing ill- onditioned weights ould result in poor disturban e
reje tion. Hen e the bounds give the designer a feel for how the weights af-
fe t the losed loop performan e. It be omes obvious, by examination of the
bounds, that the notions of lassi al loop-shaping readily arry through.
The designer an not usually augment one of the singular values of the open
loop plant (as a fun tion of frequen y) with a diagonal pre- ompensator while
leaving the other singular values un hanged. To over ome this problem the
open loop plant an be augmented with a full blo k pre- ompensator.
The singular value de omposition of the plant as a fun tion of frequen y
an be written as G(j!) = U (j!)(j!)V (j!) . If ea h element of V (j!) is
approximated with a stable minimum phase transfer fun tion to give V^ then,
G(j!)V^ (j!) ' U (j!)(j!). Hen e a diagonal weight an now be designed that
augments ea h singular value of (j!) dire tly. Note that the same an be done
when designing the post- ompensator. This method provides great exibility

69
to the designer in terms of understanding how the hoi e of weights ae ts the
a hievable performan e. Sele ting the weights in su h a way, does not ae t the
robustness of the design, as the plant is not inverted. The resulting ontroller
is given by V^ W1 K1 W2 .
As an be seen disturban e and noise reje tion, output de oupling and
tra king an easily be in orporated in the loop-shaping methodology. What has
not been dis ussed is translating the time response requirements into frequen y
response requirements. Time response requirements are spe ied in terms of
overshoot Mp , settling time ts and rise time tr with respe t to applying a step
to the referen e of the losed loop.
These requirements are set by shaping the loop gain near ross-over and
hoosing the bandwidth of the losed loop. What we must rst understand
is what kind of information we an extra t from the frequen y response of
a stable system. For example when looking at a singular value plot of the
output sensitivity of a system (stable transfer matrix) one an easily see, at a
parti ular frequen y, what the maximum gain is. So requirements of the type
that 0 (j!)  0:1 for ! < 0:1 rad/s an easily be in luded in the design
pro edure by for ing  (Lo)  11 at frequen ies smaller than 0:1 rad/s.
The Fourier series of a square wave with period
2 is given by:
!0
N
4X 1
u(t) = sin(2n 1)!0 t:
 n=1 2n 1
For a reasonably a urate representation of a square wave it is su ient to take
N = 6. Hen e a square wave of frequen y !0 = 1 rad/s an be onsidered to
ontain frequen ies ranging from 1 to 11 rad/s. In reality we an not a hieve a
perfe t square wave as the high frequen y omponents will be ltered out. If
we insert u(t) at the input of the losed loop then the output y(t) will be
N
4X 1
y(t) = jT [j (2n 1)!0j sinf(2n 1)!0 t + 6 To[j (2n 1)!0 g:
 n=1 2n 1 o
To follow a square wave of frequen y !0 (this frequen y is related to ts ) we must
make the output sensitivity su iently small over frequen ies !0 ! 11!0, hen e
ontrol the gain of To and get it as lose to unity as possible. This an be done
by in reasing the loop gain in this frequen y range. To a hieve this it might be
ne essary to in rease the bandwidth of the system (the bandwidth is related
to tr ). We must also make sure that the system is su iently well damped,
and therefore ontrol the phase of Lo (the phase of Lo is related to Mp and the
phase of To). This may mean de reasing the bandwidth due to a tuator and
sensor limitations. It ould also mean de reasing the phase lags introdu ed by
the weights at ross-over.
Even though it is not straightforward to translate time response require-
ments into the frequen y domain there are general trends that an be followed.
The rise time and overshoot are related to the damping of the system. The
less damped the system the smaller the rise time and the greater the overshoot.

70
For a desired damping ratio < 1, whi h is usually the ase for air raft, rise time
depends very little on damping. A well damped losed loop is a hieved by mak-
ing sure that the roll-o of the augmented plant singular values at rossover is

typi ally smaller than 40 dB

de
. Bode's phase-gain relationship and its gener-

alisation to the multivariable ase [61, 266 illustrates how the roll-o of (L)
is related to the phase of the loop gain and hen e the overshoot. Rise time is
ontrolled by setting the bandwidth of the system. A fast system orresponds
to a system with a small rise time and in most ases a small settling time as
well (see pp. 126-131 in [82).

7.6 Design Cy le
The typi al design y le given a plant G is as follows:
1. S ale G. The open loop plant must be s aled a ording to the desired
output de oupling and a tuator usage. This is be ause the open loop
singular values an not be asso iated with any one input or output (see
Chapter 1 in [215 and p. 42 in [120). A badly s aled plant is equivalent
to a badly formulated problem.

2. Choose the weights W1 and W2 . Integrators (or near integrators in the

ase of rate following) are pla ed in W1 to boost the low frequen y gain.
This ensures a zero steady state error if we are tra king an attitude,
disturban e reje tion and output de oupling/tra king. To in rease ro-
bust stability, hen e de rease phase lag at ross-over (i.e. slope of aug-
mented plant singular values), a proportional gain is added to the pre-
ompensator. The value of the gain (position of the resulting zero) is a
trade-o between speed of response (moves the integrator open loop pole
away from the origin in losed loop) and robustness. The bandwidth is
made as high as possible within the a tuator and sensor apabilities, i.e.
robust stability onsiderations. If the open loop is unstable are must be
taken not to make the losed loop too fast so that disturban e reje tion
leads to input saturation hen e loss of ontrollability. W2 is hosen for
noise reje tion, hen e it ontains low pass lters.

The over-all design strategy is to make the loop as fast as possible within
the limitations of the plant to use the a tuators to their limits for dis-
turban e reje tion. Open loop pre-lters are then designed to satisfy
handling quality requirements. This is based on the fa t that the distur-
ban e reje tion problem is entirely de oupled from the nominal tra king
problem (see [247 and referen es therein).

3. Choose the position of the ontroller. The ontroller an be implemented

in three ways. Pla ed in the forward path gives a faster response at the
expense of overshoots be ause all the ontroller dynami s are ex ited di-
re tly by the referen es. Also any right half plane (RHP) zeros of the
ontroller are also RHP zeros of To . An example of implementing the

71
 ontroller in the forward path is given in Figure 7.11. Pla ing the on-
troller in the feedba k path leads to a slower more damped response but
any RHP poles of the ontroller lead to RHP zeros of To. The ontroller
ould also be implemented in the observer form as proposed in [247 (see
pp. 72-78). This onstitutes the optimal way of introdu ing the ontroller
into the loop.

4. Design the ommand pre-lters. The pre-lters are designed to meet

the handling quality requirements. Performan e is limited by losed loop
non-minimum phase zeros (RHP zeros).

5. Perform time simulations and analysis to prove robust performan e. H1

loop-shaping readily provides robust stability. We an a hieve nominal
performan e but must test for robust performan e.

7.7 Two Degrees-of-Freedom Design Pro edure

The two degree-of-freedom (DOF) design pro edure as introdu ed in [117, 154
guarantees robust performan e with respe t to an ideal step response model.
Figure 7.4 illustrates the blo k diagram of the two DOF setup. The losed
loop response from the referen e signals to the plant outputs follows that of
a spe ied model Tr . The ontroller K is partitioned as K = [K1 K2 where
K1 is the pre-lter and K2 is the feedba k ontroller. The inner feedba k
ontroller K2 is used to meet the robust stability requirements while the pre-
lter K1 optimises the overall system to the ommand input. The use of the
step response model (SRM) is to ensure that

(I Gs K2 ) 1 Gs K1 Tr 1   2; (7.6)

where  is the model-mat hing parameter. From Equation (7.6) it is obvious

that as  in reases (I Gs K2 ) 1 Gs K1 ! To. By setting  equal to zero the
two DOF setup redu es to the one DOF problem des ribed earlier in Se tion
7.2.

w2 w1
r
- I - K1 -+ ?- W1 - G - ?- - I -z
+6 6
z1 y
K2 
- Tr

Figure 7.4: Two degrees-of-freedom onguration

The design y le, given a plant G with no dire t feed-through, is as follows:

72
1. Sele t a pre- ompensator W1 a ording to the guidelines given in Se tions
7.5 and 7.6. Note that in the two DOF setup W2 is usually a onstant
matrix.

2. Sele t a desired losed-loop transfer fun tion Tr between the ommands

and ontrolled outputs.

3. Set the s alar parameter  to a small value greater than 1; something in

the range 1 to 3 will usually su e.

4. Form the generalised plant P . In Figure 7.5 the signals, with respe t to
those in Figure 7.4, are: u the ontrol variables (the input to the shaped
plant GW1 ), v the measured variables (r; y ), w the exogenous signals
(r; w1 ; w2 ) and z the error signals (u; y; z ).

w - -z
- P
u v

K 
Figure 7.5: General ontrol onguration

The state spa e representation of P is given by :

2 3
As 0 0 H Bs
6 0 Ar Br 0 0 7
6 7
6 7
6 0 0 0 0 I 7
P =6
6 Cs 0 0 I 0 7
7 (7.7)
6
6 I
6
2 Cr 0 I 0
7
7
7
4 0 0 I 0 0 5
C 0 0 I 0
In Equation (7.7) H = ZCs  where Z is the solution to the Gener-
alised Filtering Algebrai Ri ati Equation. The reader an refer to [164,
215, 246 for more information on the Algebrai Ri ati Equations in
loop shaping synthesis. (As ; Bs ; Cs ) and (Ar ; Br ; Cr ) are the state-spa e
realisations of the shaped plant GW1 and referen e model respe tively.
5. Solve the standard H1 optimisation problem for the plant P to get
the ontroller K. The ontroller may be written in an observer form
as in [250.

6. Partition the ontroller in to a pre-lter K1 and a feedba k ontroller K2

and al ulate the s aling fa tor of the pre-lter as Sf = K1 (0) 1 K2(0).
The nal two DOF ontroller is illustrated in Figure 7.6.

73
 r - Sf - K1 -+ +- W1 - G -y
6
Controller
K2 

Figure 7.6: Two degrees-of-freedom loop-shaping ontroller

Note that the s aling fa tor Sf is lo ated in the ommand path. This
has been found to improve the nominal tra king properties of the losed
loop.

7.8 Design Example

We are going to present a design example to illustrate all the above points and
how the designer ould produ e a good or bad design.

7.8.1 Presentation of Model Used

We are going to design a longitudinal ontroller for the RCAM [145. The
inputs into the linear model G areT ; T H , the outputs z;_ V and the states
q; ; uB ; wB . The linearised plant model G and the denitions of these variables
an be found in [145. All angles are in rad and velo ities in m/s.
G has poles at 0:011j 0:126, 0:830j 1:107 and zeros at 4:338, 4:390.
We an easily distinguish between the phugoid and short period modes. The
phugoid is slower and lightly damped. The plant is open loop stable with a non-
minimum phase zero. The RHP zero exists be ause the verti al a elerometer
is lo ated behind the entre of rotation of the air raft. The physi al meaning
of a non-minimum phase zero is that the plant goes initially in the opposite
dire tion to that desired, so when the air raft pit hes up z_ is initially going to
be positive. Non-minimum phase zeros within the bandwidth of the plant limit
the a hievable performan e (see pp. 90-104 in [61).
The a tuator dynami s and loop delays are given below. The loop delay
transfer fun tion is based on a rst order Pad approximation (MATLAB -
tools ommands are used).

>> tailplane = nd2sys(1,[0.15 1);

>> engine = nd2sys(1,[1.5 1);
>> delay = nd2sys([-0.06 1,[0.06 1);

7.8.2 Design Spe i ations

The tra king requirements are as follows [145:

74
 Vair f
tr < 12 se tr < 5 se
ts < 45 se ts < 20 se
Mp < 5% for h > 305m
Mp < 30% for h < 305m

Table 7.1: Time response requirements

Note that z_ = V sin f , hen e for a small ight path angle f the limb rate
be omes z_  V f . For good disturban e reje tion a 13 m/s wind step should
not indu e a deviation in airspeed greater than 2:6 m/s for more than 15 se .
There are no ross- oupling requirements dened between V and z_ .

7.8.3 S aling
The open loop plant is s aled a ording to output de oupling and a tuator
usage. For onvenien e the units of the a tuators are onverted to degrees (d2r:
degrees to radians). It was thought that 1 degree of tailplane is analogous to
0:5 of a degree of thrust. This is ompatible with the physi al limits of the
a tuators [145. In reasing for example 0:5 to 1 will in rease the usage of the
throttle by in reasing the bandwidth of the system.

Bs = diag([d2r 0.5*d2r); % input s aling

Cs = diag([3 1); % output s aling
Similarly, in reasing the se ond entry in Cs from 1 to 2, will in rease the speed
of response of the airspeed and the de oupling with z_ . Figure 7.7 shows the
open loop s aled RCAM Cs *G*Bs .

3 4
10 10

2
10 2
10

1
10 0
10

0
10
singular values

singular values

2
10
1
10
4
10
2
10

6
3
10
10

8
4 10
10

5 10
10 2 1 0 1 2
10 2 1 0 1 2
10 10 10 10 10 10 10 10 10 10
frequency (radians/sec) frequency (radians/sec)

Figure 7.7: Cs *G*Bs Figure 7.8: Gs , Gs K1 and K1 Gs

Changing the dire tionality of the plant signi antly (i.e. in reasing the
loop gain too mu h in a dire tion of low plant gain) results in redu tion of the
a hieved robustness. Hen e weighting the throttle with a big number would

75
result in in reasing the losed loop bandwidth even though the throttle a tuator
is slow resulting in poor stability.

7.8.4 Choi e of Weights

To ensure that a diagonal weight W1 augments ea h of the singular values of
the s aled plant independently a weight V~ was designed as des ribed in Se tion
7.5. The total pre- ompensator be omes V~ W1 . V~ used in this design has no
dynami s. In the MATLAB ode V~ is the variable preW_V.
preW_V =
9.8313e-01 -1.0565e-01
-1.0565e-01 -9.8313e-01
Integrators are added inW1 to boost the low frequen y gain. Having just an
integrator redu es the robustness as90 of phase is added at ross-over. Hen e
a proportional matrix gain is added to W1 . The position of the zeroes is a
trade-o between speed of response and robustness. The post- ompensator W2
is designed for noise reje tion. Again the loser to ross-over that the singular
values are rolled-o at, the bigger the redu tion of the a hieved robustness.
The weights hosen are:

w1 = nd2sys(0.25*[3 1,[3 0);

w2 = nd2sys([3 1,[3 0);
preW = daug(w1,w2);
w = nd2sys(1,[0.2 1);
postW = daug(w,w);
The bandwidth in ea h loop is pushed up as high as possible subje t to a
Gs is shown in Figure 7.8 (solid
desired level of robustness. The weighted plant
K1 does not alter the desired loop shape
line). Figure 7.8 also illustrates how
too mu h. Hen e shaping G open loop is equivalent to shaping both Gs K1
(dotted) and K1 Gs (dash-dot). Note that the singular values of Gs K1 and
K1 Gs are virtually the same.
The augmented plant has 12 states. The ontroller was synthesised using

>> [sysK,emax = n fsyn(sys,1.1);

>> emax
emax =
3.3307e-01
>> emargin(sys,sysK)
ans =
3.0560e-01
Hen e the a hieved  = 0:31. The ontroller sysK was model redu ed to 7
states. Therefore the overall ontroller V~ W1 K1 W2 has 11 states. Most of
the poles in K1 are lo ated around the bandwidth. The singular values of
the model redu ed ontroller sysK2 (dashed) and the original ontroller sysK
(solid) ontroller are shown in Figure 7.9. The dieren e an hardly be seen.

76
 sysK2 gives an  = 0:30. The equations in Se tion 7.3 give a feel for the
magnitude of the a hieved gain and phase margins.
The ontroller was implemented in the forward path. The pre-lters were
hosen to be rst order lags. The singular values of So (solid) and To (dot-
ted) are shown in Figure 7.10. A -analysis ould be arried out, as in p.
3-36 [18, to he k robust performan e. Note that the large positive area under
the So urve is due to the RHP zero (see the waterbed ee t pp. 97-103 in [61).

1 1
10 10

0
10

1
10
0
10

singular values
singular values

2
10

3
10

1
10
4
10

5
10

2 6
10 10 2 1 0 1
2 1 0 1 2
10 10 10 10 10 10 10 10 10
frequency (radians/sec) frequency (radians/sec)

Figure 7.9: sysK and sysK2 Figure 7.10: To and So

7.8.5 Time Responses

The SIMULINK blo k diagram of the linear model is shown in Figure 7.11.

1 u
z_dot
3.5s+1

Csc
sysK2 W1 Bsc delays actuators RCAM y
+
1
V
7.5s+1

W2 Csc

Figure 7.11: SIMULINK blo k diagram of the linear model

Figure 7.12 shows the response to a ommand on z_ (solid) at t = 1 s. Note

the initial undershoot due to the non-minimum phase zero. Figure 7.13 illus-
trates the response to an airspeed ommand (dash-dot). Figure 7.14 illustrates
the reje tion of a wind-shear of 13 m/s.

7.8.6 Two Degrees-of-Freedom Design

Having designed a weighting fun tion W1 that provides good disturban e reje -
tion, the design spe i ations in Table 7.1 an be in luded dire tly in the design
pro edure using a two DOF approa h. The user-dened step response model,

77
 Step on z_dot Actuator usage Step on V_air Actuator usage
0.2 0.5 1.2 0.35

0.4 0.3
0 1

tailplane (solid) and thrust (dashdot) in degrees

tailplane (solid) and thrust (dashdot) in degrees

z_dot (solid) and V_air (dashdot) in m/s

z_dot (solid) and V_air (dashdot) in m/s

0.3 0.25
0.2 0.8

0.2 0.2
0.4 0.6

0.1 0.15

0.6 0.4
0 0.1

0.8 0.2
0.1 0.05

1 0
0.2 0

1.2 0.3 0.2 0.05

0 20 40 60 0 20 40 60 0 20 40 60 0 20 40 60
time (s) time (s) time (s) time (s)

Figure 7.12: Step on z_ ommand Figure 7.13: Step on Vair ommand

Tr in Figure 7.4, is usually diagonal, emphasising maximum output de oupling

and exhibiting ideal handling qualities.

>> z_dot_model = nd2sys(0.5^2,[1 2*0.7*0.5 0.5^2);

>> V_air_model = nd2sys(0.3^2,[1 2*0.7*0.3 0.3^2);
>> T_r = daug(z_dot_model,V_air_model);

As des ribed in the design y le  > 1. The nal value was  = 1:5. A few
iterations were required (bearing in mind robust performan e) before arriving
to this hoi e. The generalised plant was formed from Equation (7.7) and a
slightly suboptimal ontroller was obtained using standard H1 optimisation
routines [18. The degradation of the stability margin ( ) as  in reases is
shown in Table 7.2. It is evident that the better the model-mat hing the less
robust is the design we an a hieve. Balan ed residualisation (see pp. 449-454

 1.1 1.2 1.3 1.4 1.5 1.6 1.7

4.15 4.30 4.44 4.59 4.73 4.91 5.06

Table 7.2: Stability margin as a fun tion of 

in [215) was used to redu e the ontroller to 8 states. The ontroller was im-
plemented as in Figure 7.6. Figure 7.15 shows the output response to a unit
step input on z_ . Figures 7.16 and 7.17 illustrate the responses to an airspeed
demand and a 13 m/s wind shear respe tively.

By omparing the output oupling in Figures 7.12, 7.13 and 7.15, 7.16 it is
evident that the two DOF s heme gives good performan e without signi ant
deterioration of the losed loop robustness properties.

It an be dedu ed that all the requirements are met. The interested reader
is en ouraged to go through the example and hange the s aling and weighting
fun tions to obtain a feel of how the dierent parameters inuen e the design.

78
 Disturbance on z_dot Actuator usage Step on z_dot Actuator usage
15 10 0.4

0 0.3

tailplane (solid) and thrust (dashdot) in degrees

z_dot (solid), V_air (dashdot), SRM (dot) in m/s

tailplane (solid) and thrust (dashdot) in degrees

10 5
z_dot (solid) and V_air (dashdot) in m/s

0.2 0.2

5 0
0.4 0.1

0.6 0
0 5

0.8 0.1

5 10
1 0.2

10 15 0.3
0 20 40 60 0 20 40 60 0 20 40 60 0 20 40 60
time (s) time (s) time (s) time (s)

Figure 7.14: 1DOF wind-shear Figure 7.15: 2DOF z_ demand

7.9 Limitations of the Method and Ideal Plant

Some plants have features that restri t the a hievable performan e (see pp.
143-153 in [266 and Chapters 5, 6 in [215). Su h limitations are for example
RHP poles outside the bandwidth, RHP zeros within the bandwidth and ill-
onditioning. These limitations are design method independent.

Step on V_air Actuator usage Disturbance on z_dot Actuator usage

0.4 15 10

0.35
1
z_dot (solid), V_air (dashdot), SRM (dot) in m/s

z_dot (solid), V_air (dashdot), SRM (dot) in m/s

tailplane (solid) and thrust (dashdot) in degrees

tailplane (solid) and thrust (dashdot) in degrees

10 5
0.3
0.8

0.25
5 0
0.6

0.2

0.4
0 5
0.15

0.2
0.1
5 10
0
0.05

0.2 0 10 15
0 20 40 60 0 20 40 60 0 20 40 60 0 20 40 60
time (s) time (s) time (s) time (s)

Figure 7.16: 2DOF Vair demand Figure 7.17: 2DOF wind shear

H1 ontroller design te hniques are frequen y domain based methods. This

is be ause robustness issues are more easily addressed in the frequen y do-
main. The potential weakness of H1 loop-shaping is that there is sometimes
di ulty in translating time response requirements to the desired loop shape.
This di ulty an be over ome by pushing for the highest possible losed loop
bandwidth. Hen e the designer aims for a bandwidth higher than that required
to satisfy the handling quality requirements, subje t to obtaining reasonable
robustness.
The design problem be omes harder when the plant is open loop unstable
in whi h ase a high bandwidth ould lead to input saturation and loss of on-
trollability during disturban e reje tion. In su h ir umstan es the bandwidth
must be lowered. De oupling be omes a more di ult task, parti ularly if the
spe i ations for ea h loop vary signi antly.

79
 The ideal plant for ontrolling with an H1 loop-shaping design would be
a plant that has similar properties in all loops. By similar properties we mean
equally fast and powerful a tuators and sensors and not too ill- onditioned.
Hen e the ross-over frequen y for all singular values an be made the same.
When using lassi al ontrol the designer designs the ontroller dire tly.
This is not the ase when using H1 design te hniques as the ontroller is
the produ t of an optimisation and hen e the designer has to set-up the ost
fun tion to be minimised. There is an evident transfer of tasks. As H1 loop-
shaping provides robustness to a very general lass of perturbed plants the
designer has only got to worry about translating the performan e spe i ations
to the desired loop shape.
Other examples, tutorials, of designing loop-shaping ontrollers an be
found in [18, 120, 215, 247.

80
8. -Synthesis

Samir Bennani1 , Gertjan Looye 1

and Carsten S herer 2

8.1 Introdu tion

This hapter gives some ba kground theory on the Stru tured Singular Value,
, and provides a ight ontrol design example motivating and demonstrating
the ne essary steps to arry out a -synthesis design. The issue addressed in
the example illustrates the inherent ontrol paradigm that -synthesis partially
solves.
Fundamentally,  addresses the problem of retaining a desired performan e
level in the fa e of un ertainties, whi h is alled the robust performan e prob-
lem. For SISO systems, this is automati ally a hieved when the system has
guaranteed robust stability and nominal performan e. This does not hold in
the MIMO ase and in this respe t, the - on ept is a tool to address the
multivariable robust performan e problem.
An important by-produ t of the method is that it rises modeling issues in the
most general sense, i.e. that we mean system modeling, spe i ation modeling,
un ertainty modeling, open loop or losed loop modeling and their validations
are all issues whi h appear on e a designer is fa ed with . An attempt in
ta kling and predi ting the real world an be done only by formal tools, and
this is where  is intended to be used. The singular value loop shape paradigm
as presented in [64 was a great leap forward in formalizing robust multivariable
ontrol theory. This resulted in progress towards H1 optimal ontrol, for
whi h omputable ee tive solutions are presented in [62, 62, 63. The lassi al
multivariable feedba k problem is illustrated in gure 8.1. Usually, the plant G
is an element of a set of plants given by G~ . We shall onsider that ea h system
in the set G~ is linear, nite dimensional and a time invariant system whi h
an be represented by a transfer fun tion G(s) . The overall design problem
onsists of three subproblems. The Robust Stabilization problem (RS) is to
nd a ompensator K whi h makes the feedba k loop in gure 8.1 internally
stable for all possible plants G~ . The se ond problem, the Robust Performan e
problem (RP), whi h is mu h harder to a hieve, requires the ompensator K
to make the losed loop system respond well under various external signals.

1
Fa ulty of Aerospa e Engineering, Stability and Control Group, Delft University of Te h-
nology, Kluyverweg 1, 2629 HS Delft, The Netherlands
E-mail: s.bennanilr.tudelft.nl
2
Me hani al Engineering Systems and Control Group, Delft University of Te hnology,
Mekelweg 2, 2628 CD Delft, The Netherlands.

81
 n d
e u ~
r K G y
-

Figure 8.1: Classi al feedba k onguration

This means that for all plants in G~ , the plant outputs y(t) a urately tra k
external ommands r(t) , even in presen e of disturban es ( d(t) ) and sensor
noise ( n(t) ). A third problem alled the Control Eort minimization problem
is a onstraint imposed on the ompensator su h that the ontrol signals u(t)
and/or other ontrol dependent signals remain within appli able limits.

It has been remarked in [64, 65, 58 that the singular value on ept leads to
onservative robust performan e predi tions. Therefore, the stru tured singu-
lar value  has been proposed as a more rened robust performan e indi ator.
In Beyond Singular Values and Loop Shapes [227 by Stein and Doyle the
singular value loop shaping as a paradigm for multivariable feedba k system
design in the arrangement as shown in gure 8.1 has been revisited. The main
on lusion drawn was that singular values within the lassi al design framework
are ee tive in addressing the performan e robustness problem whenever the
problem's design spe i ations are spatially round, but that it an be arbitrar-
ily onservative otherwise. The origin of the problem lies in that onditions for
robust performan e based on singular values are not tight (su ient, but not
ne essary) and an severely overstate a tual requirements. The onservatism of
the singular value loop shape paradigm in the lassi al framework ame from a
too narrow denition and representation for a system. Furthermore, a general
tight performan e spe i ation pro edure is la king. Finally, the stability anal-
ysis and synthesis tools were not addressing the fa t that perturbations arising
in the system are stru tured. The onservatism introdu ed when using singular
values an be surmounted by using the Stru tured Singular Value (SSV)  as
a tighter multivariable generalization of the stability margin. It will be shown
that  naturally arises from the stability analysis of a general lass of systems
alled Linear Fra tional Transformations (LFT's). Naturally, in the sense that
the existen e of LFT's automati ally leads to the formulation of the robust
performan e problem. General, in the sense that LFT's are both suitable for
the analysis and the synthesis problem.
Using LFT's to model sets of systems and the ontrol obje tives in mind,
the robust MIMO design problem is formalized by spe ifying, the plant set
G~ over whi h the obje tives must be a hieved and the pre ise mathemati al
statements for the performan e and ontrol eort obje tives.
This will be illustrated on a simplied ight ontrol problem that we de-
s ribe rst. The design plant is a linear model of the longitudinal short period
dynami s of a Cessna Citation 500 in landing onguration. The model states

82
are the pit h rate q and the angle of atta k . The state spa e representation
of the model dynami s is given as:
      
q_ = Mq M q + ME E (8.1)
_ 1 Z  ZE
The input is the elevator dee tion E . The ontrol obje tive is to design a
pit h rate ontroller, su h that the losed loop response mat hes the handling
quality model Hid (s) = qq ((ss)) = s+1
1:5 .
:5 From robustness onsiderations we
have to ensure that the system works well in the fa e of un ertain state spa e
entries, alled the stability and ontrol derivatives, for trim speed variations
up to  10 m/s. During a full pit h ommand manoeuvre the angle of atta k
is limited to jj < 20 deg and the elevator dee tion and dee tion rate to
(jE j; j_E j) < (10; 30) [deg, deg/s.
The mathemati al formulation of the performan e spe i ations in the on-
trol problem and the model set over whi h these spe i ations have to hold an
be done by using linear fra tional transformations and norm bounds. The ad-
vantage of the LFT formulation is that it gives a ommon base for un ertainty
modelling, stability and performan e analysis of perturbed systems ( alled the
analysis problem) and nally for ontroller synthesis (our synthesis problem).
Ea h of these three steps will be su essively illustrated by an appli ation on
the air raft example. To illustrate the pra ti allity of  as mature design tool
we on lude the example with a trade-o study, where the performan e and
the robustness in the problem are gradually hanged.

8.2 Linear Fra tional Transformations (LFT's)

Denote M as a 2  2 blo k-stru tured matrix:
v1 = M11 r1 + M12 r2
v2 = M21 r1 + M22 r2
together with matrix
relating v2 to r2 as v2 =
r2 . Closing the lower
hannels of M with
gives a Lower Linear Fra tional Transformation of M
by
:
h i
v1 = M11 + M12
(I M22
) 1 M21 r1 = Fl (M;
)r1
here l indi ates that the lower hannels of M have been losed with
. In
the same way we an lose the upper hannels of M with some appropriately
dimensioned matrix
that relates r1 to v1 in the following manner: r1 =
v1 .
The upper LFT is given by:
h i
v2 = M22 + M21
(I M11
) 1 M12 r2 = Fu (M;
)r2
Many ontrol problems t within this representation. A well known example
is the input-output mapping of a linear system, y = G(s)u. It an be expressed

83
 in terms of state spa e data as an LFT system. It easy to see that G(s) =
C (sI A) 1 + D an be rewritten as
  
G(s) = Fu A B ;I
C D s
As we shall see this framework is parti ulary suitable to arry out parametri
un ertainty modelling.
This is illustrated on the air raft problem where due to hanging operating
onditions the state spa e entries of the nominal model (equation 8.1) vary
substantially. In table 8.1 the nominal values of the elements and the max-
imum relative variations an be found. Drawing the system dynami s in a

parameter value mult. pert.

M -1.4796 0.20
ME -6.7679 0.20
Mq -1.5773 0.20
Z -0.7441 0.20
ZE -0.0900 0.20

Table 8.1: State spa e elements and perturbations for the design example

z1
.
1 Z
1

+ + Z
w1

z2
..
Z
w2
2 Z
E
. 5
z3
5 M
w3
M

4
z4
M q z w
. + w4
A/C
q Mq q
+
z5
M
E
3
M
w5 q

Figure 8.2: Blo k diagram of exam- Figure 8.3: Representation of ex-

ple system with perturbations ample system in LFT-form

blo k diagram, we obtain insight in how the perturbations ae t the model, see
gure 8.2. The perturbations in the table are the maximum absolute hanges
the parameters an undergo whi h are normalized to i by introdu ing s alings.
We may hange any of these 's arbitrarily within the given bounds. The model
parametrized in the 's ree ts a set of models. To derive an LFT representa-
tion, the invariant part of the model and the un ertain elements (the delta's)
are separated. This pro ess is known as pulling out the deltas. All un er-
tain elements 1 ... 5 are diagonally augmented in the perturbation matrix

= diag(1 ... 5 ). In gure 8.2 the signals in and from the delta's have been
ut; the signals (z1 ...z5 ) be ome the outputs from the onstant part and inputs
of
, while the opposite holds for the signals (w1 ...w5 ). We an now read o

84
all signal relations given by the mapping and build the following matrix:

 q w1 w2 w3 w4 w5 E
_ Z 1 Z ZE 0 0 0 ZE
q_ M Mq 0 0 M Mq ME ME
z1
Z 0 0 0 0 0 0 0
z2 0 0 0 0 0 0 0 Z
E
z3
M 0 0 0 0 0 0 0
z4 0 Mq 0 0 0 0 0 0
z5 0 0 0 0 0 0 0 M
E
 1 0 0 0 0 0 0 0
q 0 1 0 0 0 0 0 0
The so obtained matrix losed with the blo k diagonal matrix
as shown
in gure 8.3 provides the required LFT formulation of the un ertain air raft
dynami s overing all possible parameter variations. Noti e that the para-
metri un ertainty modeling pro ess reveals that un ertainty that is unstru -
tured at parameter level ( omponent level) be omes stru tured at system level
A/C).
( Another possible way to apture a set of models is given for the
a tuator. The elevator position is hanged via an a tuator having rst order
dynami s
15 .
Ga t (s) = s+15 The devi e is assumed to give position errors up
to 20% in a frequen y range up to 1 rad/s, while at higher frequen ies this
may be even more. The variation in the position error along the frequen y
is represented by a rst order transfer fun tion
1+1 . The
Wpert (s) = 0:20 s=s=40+1
model set overed by the un ertain a tuator dynami s is given by G ~ a t =
fGa t (1 + 6 (s)Wpert ) : 6 (s) stable k6 (s) k1  1g. The weighting fun tion
is used to normalize the unknown perturbation 6 and at any frequen y ! the
magnitude of Wpert represents the relative un ertainty level in the a tuator
model. This type of modeling is alled multipli ative un ertainty modeling.
It is unstru tured at omponent level (the a tuator). On e again, we shall
see how unstru tured un ertainty at omponent level be omes stru tured at
system level, when inter onne ting these omponents.

8.3 The Extended Design Framework

The performan e robustness problem is addressed in a general framework for
system design, whi h onsists of a general problem des ription in terms of LFT
systems, some key analysis results, a suitable measure of the magnitude for
matrix transfer fun tions, the stru tured singular value , and some ontroller
synthesis results. An important remark is that all elements onsidered in this
design framework have a pra ti al software implementation in order to be useful
for the engineering world. This issue is provided by the ex ellent and reliable
software of [18. Furthermore, the pra ti ing engineer needs good tutorials to
keep tra k of the theoreti al advan es in the eld, these are nowadays ri hly
provided in referen es as [18, 189, 266, 61 and many others. The various
elements of that framework are des ribed next.

85
8.3.1 General problem des ription
Using the LFT representation the lassi al multivariable ontrol problem as
shown in gure 8.1 an be transformed into a more versatile form. In this way
any performan e obje tive from the a tual inter onne tion and its ee t within
other system loops is derivable.

w z
d P e
u y
K

Figure 8.4: The general problem des ription

We re ognize for any un ertain losed loop system the three basi ompo-
nents:

1. general system P (our problem data)

2. the ontroller K (possibly still to be designed)

3. the un ertain elements
(belonging to a pre-spe ied set).

All un ertain elements have been pulled out of the system and pla ed in the

-blo k. For synthesis and analysis the only thing we have to know is that
the
-blo k is stable and norm bounded: jj
jj1  1. The always returning
subdivision for the general system P onsists of re ognizing three pairs of
input-output variables. The rst one (u(t); y (t)) onsist of the ontrol and
measurement variables. Then we have (d(t); e(t)), the disturban e and error
signals whi h onstitute the generalized performan e variables and nally the
third pair (w(t); z (t)) for the perturbation signals whi h are onne ted ba k
into the system through a norm bounded perturbation
.
The design problem is to nd a ompensator K internally stabilizing the
general system P while keeping the matrix transfer fun tion between d and
e appropriately small for the whole set of allowable perturbations
. In the
transformation pro ess, from the lassi al setup into the more general setup,
any un ertainty arising at system omponent (a tuator, plant sensors et .)
level be omes automati ally stru tured at the level of the generalized system
P. Furthermore, the so obtained generalized problem des ription as given in
gure 8.4 is suitable for the synthesis as well as for the analysis problem, and
has potential for expansion due to its general stru ture.

8.3.2 Analysis results (Doyle 1984)

From the general system representation as shown in gure 8.4, a non onserva-
tive, ne essary and su ient ondition for robust performan e an be derived.

86
 0
0 p

w z w z
M ( P, K ) e
M ( P, K )
d d e

Figure 8.5: Closed loop system M Figure 8.6: Stability onguration

for Robust Performan e

Suppose the stabilizing ontroller K an a ept the feedba k loop as shown

in gure 8.4 to get the losed loop system M (P; K ) = Fl (P; K ), shown in g-
ure 8.5. The generalized losed loop system is a 2  2 blo k-stru tured transfer
fun tion matrix M (P; K )(s) whi h together with the
operator in a feedba k
arrangement, forms the basi obje ts on whi h the system analysis problem is
based.
Under the ondition the system M (P; K ) is nominally stabilized by K the
following results apply:

Theorem 8.1 General Analysis Theorem Doyle (1984)

1. Nominal performan e is satised if and only if

 (M22 (j!)) < 1 8! (8.2)

2. Stability is robust if and only if

 (M11 (j!)) < 1 8! (8.3)

3. Performan e is robust if and only if

(M (j!)) < 1 8! (8.4)

The third result represents the MIMO extension of the robust performan e
problem, providing with ne essary and su ient onditions. It is established
by starting with the denition that performan e is robust if and only if the
transfer fun tion matrix from d !e given by Fu (M;
), remains in an H1
norm sense bounded by unity - that is, if and only if


 M22 + M21(I
M11 ) 1
M12 < 1 8 !; k
k1  1 (8.5)

This norm bound is also a ne essary and su ient ondition for the system
M in gure 8.6 to remain stable if we onne t a se ond norm-bounded per-
turbation, say
p (s) a ross the e ! d terminals In this respe t, robust

87
performan e is equivalent to robust stability in the fa e of two perturbations,

and
p , onne ted around the system M in the blo k-diagonal stru tured
arrangement shown in gure 8.6. The system M is robustly stable if and only
if the fun tion det(I diag (
;
p )M (j!)) remains non-zero for all !.

8.3.3 The Stru tured Singular Value

The observation to view performan e robustness as a stability test brought the
fun tion (:) dened in equation 8.4. The fun tion  for omplex matri es M
is the following:

Denition 8.2 The Stru tured Singular Value 

2 8  93 1

4min <  det(I 
M ) = 0 =
(M ) =  for some
= diag[
1 ;
2 ; : : : ;
m 5 : (8.6)
: 
with  (
i )  1 8 i ;

In words,  is the re ipro al of the smallest value of s alar  whi h makes the matrix
I 
M singular for some
in a blo k-diagonal set. If no su h  exists,  is
taken to be zero. This denition redu es to the onventional maximal singular value
in absen e of stru ture ( i.e. when the number of blo ks, m , in
is one ). For this
reason,  has been alled the stru tured singular value.

The denition 8.6 is not limited to 2 2 blo k stru tures. It an be used

to test stability with respe t to any number of diagonal blo ks: in that ase
robust stability is satised if and only if


pert (M11 (j!)) < 1 8! (8.7)

for a given blo k-stru tured un ertainty from the set
pert . In this way it is
possible to establish robust stability with respe t to plants ae ted by several
stru tured perturbations while tting in the robust performan e paradigm.
Denition 8.6 also extends to real-valued perturbations redu ing many para-
metri system analysis problems to - al ulations. More generally still, the
stru tured singular value on ept (not value) extends to time varying systems.
The al ulations required for these extended ases expressed in Linear Matrix
Inequalities, ontinue to impose substantial hallenges even with the tremen-
dous evolution in the eld of onvex optimization.

8.3.4 Numeri s for the stru tured singular value

In general exa t omputation of the stru tured singular value is not possible.
Therefore, we work with approximations via the upper and lower bounds of .
For a omplete tutorial on the stru tured singular value  and the involved
numeri s we refer to [188, 18. The al ulation of the stru tured singular
value () relies always on the parti ular hoi e of the un ertainty stru ture

and generalizes matrix measures as the singular value and the spe tral ra-
dius. Based on this generalization omputable bounds an be given and re-
ned. Dene the following perturbation stru tures
= fdiag[1Im ; 2;
3 :
88
1 ; 3 2 C;
3 2 Cnn ; g. The set B
is the sub-set of
for whi h holds B
=
f
2
: (
)  1g. We an also dene Q
, Q
= f
:
2
;

= I g
I I

It an be shown that Q
2
,
Q 2
,  (Q
) =  (
) =  (Q
) Asso i-

 
ated with the set
, dene the set of s aling matri es D given by:

D = diag[D1; d2 ; d3Ik : D1 2 Cmm ; d1 ; d2 2 IR + ; D1 = D1 > 0;

I

Note that where the diagonal blo k of
is full D is a s alar, and vi e versa. It
an be seen dire tly that for ea h D 2 D and
2
holds, D
=
D. In
the ase
2 C
nn it is easy to see that  (M ) =  (M ). Sin e the perturbation
I

is bounded we have 
(M )    (M ). However, this bound is not of pra ti al
use sin e the gap between  and the   an be arbitrarily large. On the other
hand when
= 1 Im , with 1 2 C then (M ) = (M ), the spe tral radius of
I

M . Using the transformations D and Q on M the bounds an be rened to:

max (QM )  (M )  inf  (DMD 1 )
Q2Q
D 2D
(8.8)

In fa t the left inequality is an equality, but not useful as su h, sin e the

optimization over Q is not onvex; it shows lots of lo al minima and maxima.
More useful is the right inequality (whi h in a limited number of ases is also
an equality, but in nearly all ases very tight), sin e the optimization over D
is onvex. Furthermore, it is an upper bound and therefore safe.

The perturbations onsidered, were omplex matri es or s alars. However,

in the ase of the parametri un ertainties in the air raft, the perturbations
are real ( 2 IR ). In a ase like this, we would like to know, given the system
with un ertain parameters, the smallest possible ombination of perturbations,
that auses the system to be ome unstable. It is obvious that the stru tured
singular value for this even more onstrained set of perturbations (un ertainties
are only allowed to vary along the real axis) is more di ult to determine. For 
al ulations with a mixed omplex/real perturbation set, there exist reasonably
tight upper bounds by nding optimal s aling matri es (D-G s ales), for more
details the reader is referred to [266, 18.

8.3.5 Setting up the design problem

For the analysis of the ight ontrol example, we rst have to spe ify the overall
ontrol ar hite ture then translate the design spe i ations into mathemati al
obje tives by weighting the signals of interest. To demonstrate the exibility of
the proposed framework we shall address the simultaneous design of a feedba k
and ommand shaping ompensator whi h is often referred to as two degrees of
freedom ontrol. Upon the hosen ontrol ar hite ture we pla e on the physi al
lo ation in the system, namely at signal level our requirements. These require-
ments are made frequen y dependent and be ome our weighting fun tions. In
doing so we end up with the situation depi ted in gure 8.7 whi h is what we
all the inter onne tion stru ture. It onsists of the air raft model G(s) with its
parametri un ertainties, the un ertain a tuator Ga t (s), the ontrollerK (s)

89
to be designed, an ideal model Hid (s) whi h we want to mat h and the per-
forman e weighting lters that pla e emphasis on the frequen y ontent and
amplitude on the signals of interest. The inter onne tion stru ture in gure 8.7
is the pi torial equivalent of the mathemati al statement of the plant set to-
gether with the ontrol spe i ations (depending on the norm we hoose). It

~ ~.
E E

1. 1 1 z6
z .. w .
5 max max 5 Wpert
qc
~ 1 w
+ 6
u Win q nom
max G 1
. 15 K q
E S
E -

Gact
q

n Wn

~ -
qe Wp H id

Figure 8.7: Inter onne tion stru ture of the example system

is often advisable to s ale the systems units appropriately. The nominal pit h
rate ommand signal is therefore normalized with respe t to the maximum ex-
pe ted ommands with the lter Win . The pit h rate ommand input q goes
through the ideal model. The dieren e between the ideal model response and
the a tual pit h rate measurement q is the tra king error. To emphasize how
large and up to what frequen y the error redu tion should o ur, a lter Wp (s),
ree ting the tra king obje tive, is pla ed on the error signal. However, tra k-
ing should not be a hieved at the ost of ex essive ontrol a tivity. Therefore
both the elevator dee tion and rate are penalized. The dee tion and rate are
weighted by the inverse of their maximum allowed values E Emax and W = 1=
W_E = 1=_Emax . When one of these weighted signals is larger than one, then
the obje tive is violated. To prevent stall during a full pit h rate ommand
we provide an angle of atta k limiting fun tion by introdu ing a performan e
spe i ation on , using the inverse of the maximum allowable value we get the
weight W = 1=max. Finally, a noise lter Wn is depi ted in gure 8.7. This
lter s ales the normalized measurement noise n as a fun tion of the frequen y.

Dis onne ting in gure 8.7 the ontroller and the un ertainties we end up
with the open loop inter onne tion stru ture P as shown gure 8.8. The re-
maining system P :~:-sign indi ates
has three pairs of inputs and outputs (the
a weighted output) These orrespond to the un ertainty hannel ( z w), the
performan e hannel given by e = [~ ~ ~
_
qe ; ~; E ; E , d = [qnom ; n and the mea-
surement/ ontrol hannel with y = [q ; q + Wn n, and u = [E . It ontains
all required problem data for design. But sin e the weighting fun tions are in
most of the ases our design parameters it is worth to start with the analysis
on basis of the hypothesis that we are in possession of a stabilizing ontroller
K (s).

90
 3
10
Z 1.... Z 5 w 1 .... w5 Wp2

(real) (real) 2
10
Z6 Wp1
(complex) Z6 1

~q 10

e (complex) Hid
~ 0

10

gain
~ q nom
E
1
10

.

~ 2
10
Werr

E n Wn

qc 3
10

(noisy) q E 4
10 2 1 0 1 2 3
10 10 10 10 10 10
frequency (rad/s)

Figure 8.8: Open loop inter onne - Figure 8.9: Weighting fun tions for
tion stru ture design example (not s aled with
Win )

8.3.6 Weighting fun tion sele tion onsiderations

Up to now, we have dened a set of models (nominal air raft model with
perturbations), de ided on the ontroller ar hite ture (two degrees of freedom,
measurements et .), and whi h performan e quantities we wish to take into
a ount (tra king q, maximum ontrol dee tions et .). The question now
arising is how do the weights have to look like if we want our losed loop
system to a hieve robust performan e?

Mu h an be learned about the system by writing down the transfer fun -

tions in the inter onne tion stru ture. For simpli ity, we will forget about the
parametri un ertainties in the air raft model and onsider only the un ertainty
at the a tuator.
Re all that the rst step is to obtain the transfer fun tions of the open loop
inter onne tion stru ture P (gure 8.8). These fun tions an be dire tly read
from the inter onne tion stru ture in gure 8.7. Denote the transfer fun tion
from E to q is Gq and from E to  as G . The 3  3 open loop inter onne tion
stru ture P is given as:
2 3 2 3
z 0 0 0 Wpert
6 q~e 7 6 Wp Gq Ga t Wp Hid Win 0 Wp Gq Ga t 72 3
6 7 6 7 w
6  7 6 7
6 ~ 7 6 W G Ga t 0 0 W G Ga t 76
6 ~ 7 6 7 6 qnom 7
6 E 7 = 6 0 0 0 WE Ga t 74 7 (8.9)
6~
6 _ E
7
7
6
6 0 0 0 W_E sGa t 7
7
n 5
6 7 6 7 E
4 q 5 4 0 Win 0 0 5
q Gq Ga t 0 Wn Gq Ga t
All expressions required for analysis of the losed loop system M = Fl (P; K )
an be obtained via the short ut:

Mij (s) = Pij + Pi3 (s)[I K (s)P33 (s) 1 K (s)P3j (s) i; j = 1; 2

91
with K (s) partitioned into [K Kf , here K (s) represents the ommand part,
given by the transfer fun tion from q to u, while Kf (s) for the feedba k task
stands for the transfer fun tion from q to u.
The input sensitivity fun tion Si is given by:

[I K (s)P33 (s) 1 = [I [K Kf [0 Gq Ga t 1
= [I Kf Gq Ga t 1 = Si
so that with the omplementary sensitivity Ti we get Si + Ti = I .
The omplete analysis system M is:

2 3T 2 3
Wpert Ti Si K Si Kf 2 3T
6 Wp 7 6 Gq Ga t Si Hid Gq Ga t Si K Gq Ga t Si Kf 7 I
6 7 6 7
diag 6 W 7 6 G Ga t Si G Ga t Si K G Ga t Si Kf 7 diag 4 Win 5
6 7 6 7
4 WE 5 4 Ga t Ti Ga t Si K Ga t Si Kf 5 Wn
W_E sGa t Ti sGa t Si K sGa t Si Kf
Sin e we try to minimize  approximated by the peak value of the s aled
singular value (  (DMD 1 ), a -upperbound), preferably to a value lower than
1, the diagonal blo ks (not ae ted by the D-s ales) must have a norm smaller
than 1. This leads to the robust stability and nominal performan e onditions:

jjWpert Ti jj1 < 1

and:
2
Wp 0

0 0 32 Hid Gq Ga tSi K Gq Ga t Si Kf 
3


6 0 W 0 0 76 G Ga tSi K G Ga t Si Kf 7 0 Win 0
4 0
0 WE 0 54 Ga tSi K Ga t Si Kf 5 0 0 Wn < 1
0 0 0 W_E sGa t Si K sGa t Si Kf
1
From the rst expressions (M11 ) it is lear that the weight Wpert for es the
loop gain to roll o: at low frequen ies jTi j  1 and Wpert will be taken 20%. At
higher frequen ies jTi j  Kf G, while Wpert will in rease (more un ertainty to
a ount for unmodeled dynami s et .). For jWpert Kf Gj < 1 (SISO in this ase),
we need at least: jKf Gj < 1=jWpert j. In this example, it will have little ee t,
sin e the plant with its a tuator model have already su ient roll-o. How-
ever, the ross-over frequen y of Wpert is used as a design parameter to limit the
bandwidth of Ti . This is important to prevent ex itation of, for example, stru -
tural modes by ontrol signals ontaining high frequen ies. A more interesting
ase is the performan e blo k (M22 ). To a hieve the tra king obje tive, we need:
jWp (Hid Gq Ga t Si K )Win j < 1. At low frequen ies jHid j = 1 and we assume
that K  Kf : jWp (I + Gq Ga t Si Kf )Win j = jWp (I Gq Ga t Kf ) 1 Win j < 1
If the loop gain is high, we have approximately: jWp (Gq Ga t Kf )
1 Win j < 1,
so that: jGq Ga t Kf j > jWp Win j. We will hoose Win onstant. Gq Ga t have
onstant gain at low frequen ies, so that we mainly inuen e the low frequen y
shape of the ontroller K via Wp . Observe that the tra king performan e
may be destroyed by the noise input via Wn : Wp Gq Ga t Si Kf Wn . Sin e
jGq Ga t Si Kf j  1 at low frequen ies, we need at least: jWp Wn j < 1. But, if

92
we give Wp high gain at low frequen ies to in rease the gain of the ontroller,
this requirement is easily violated. There are two simple solutions: in the rst
pla e we an make the noise input low where Wp is high, so that Wn gets
approximately the inverse shape of Wp ; in the se ond pla e we an feed the
noise to the performan e lter, so that Hid q (q + Wn n) is weighted. This
is very obvious: the steady state value (at ! = 0) of Wn a ounts for example
for a sensor bias. This bias may violate the performan e index, be ause this
is based on the error between the referen e and the exa t output. By applying
the se ond solution, the error is related to the same biased measurement the
ontroller re eives. (In a standard feedba k onguration the transfer fun tion
of the noise to the output is hanged from a omplementary sensitivity fun tion
to a sensitivity fun tion, whi h has low gain at low frequen ies). We an hoose
here to lower the gain of Wn if ne essary; this has a desirable ee t when we
design a ontroller for the plant without un ertainties, as will be shown later.
The terms W _ sG S K W
E a t i in and E W _ sGa t Si Kf also play an important role,
mainly at higher frequen ies. In that ase jSi j  1, sGa t = 15j!=j! + 15  15
as s ! 1, so that at least: jK j < 1=jW_E 15Win j and jKf j < 1=jW_E 15Wn j
respe tively. These weightings impose an upper-bound on the high frequen y
gain of the ontroller. In many ases the ontroller rolls o at higher frequen-
ies, so that the weights do not have a great ee t. However, in the ase of
plant perturbations or severe disturban es it is very important to penalize the
rates to prevent the ontroller from produ ing ontrol signals with rates be-
yond the physi al limits of the plant, ausing rate saturation. We will design
for ommands ( q ) up to Win = 10=57:3 rad/s. Next, two performan e weights
are hosen, to illustrate their ee t on the ontroller shape:

s=20+1 W = 1000 s=20+1

Wp1 = jW20in j s= 0:5+1 p2 jWin j s=0:01+1 (8.10)

Note, that the ross-over frequen ies are equal. This is an important onsid-
eration. In the low frequen y range there are two major parameters for the
shape of Wp : the steady state gain and the ross-over frequen y. We must be
areful to hange one at a time. If the steady state error appears to be too
large (in a simulation for example) simply in reasing the gain means that also
the ross-over frequen y in reases, leading to unintended other ee ts. Usu-
ally, if a good ross over point is found, one an try to extend the slope into
the low frequen y range. This an be seen for Wp1 and W p2 in gure 8.9.
Note that the weight attens at ! = 10 rad/s. This is useful to limit tran-
sient behaviour. The weights on the elevator dee tion and rate are hosen
1= 5710:3 rad 1 _E 1 = 5730:3 (rad/s) 1 To limit the angle of at-
as: Emax
1 1
max
57:3
ta k we hoose
max = 2jWin j = 20 . Finally, we dene the noise lter:
0
Wn = 0:0005jWinj s=2+1 = 57
s= : 01+1 0:01 :01+1 . The DC gain is hosen low taking
s= 0
:3 s=2+1
into a ount that we also have to satisfy the performan e index Wp2 .

93
8.4 -Synthesis
8.4.1 Formulation of the synthesis problem
The next step is the ontroller synthesis problem. The obje tive is to nd a
stabilizing ontroller K a hieving the desired performan e requirements for the
whole set of plants P~ .
P~ = fFu (P;
pert ) :
pert 2
pert ; k
pert k1  1g (8.11)

The denition of the -synthesis obje tive:

Denition 8.3 -synthesis :
Minimize over all stabilizing ontrollers K (s) the worst ase performan e, i.e.
the peak value of 
(M ) = 
(Fl (P; K )).
min
K (s)
k[Fl (P; K )k
< 1 (8.12)

stabilizing

with the shorthand notation of the -norm of the operator G and similarly
to the 1-norm we have kGk
= max! 
(G(j!)). The stru tured singular
value does not satisfy the denition of a norm. This notation is adopted only to
ree t the fa t that we want to measure the size of the worst ase performan e.
In order to perform al ulations we repla e 
(
) by its upper bound  (D(
)D 1 ).
Dene Dpert , the s aling set for the perturbation stru ture
pert . For Dpert 2
Dpert and
pert 2
pert it follows from the denition of the invarian e of  un-
der s aling, Dpert
pert =
pert Dpert , that the s aling set D for the augmented
perturbation set is dened as:
  
D= Dpert 0 : where Dpert 2 Dpert (8.13)
0 I
Note that theD-s ale orresponding to the performan e blo k
p -blo k is set
to one. With respe t to the s aling stru ture D 2 D 
an be obtained
from the upper bound relation (applied to some onstant matrix M ):


(M )  min (DMD 1 )
D 2D
(8.14)

When
pert onsists of F full blo ks, the set D looks like

D=f diag [d1 I; : : : ; dF I; I dj > 0g (8.15)

The elements ofD an have any phase without ae ting the value of  (DMD 1 ).
Therefore the optimization along the frequen y over D an be repla ed by an
optimization over stable minimum-phase D (s). Considering real-rational, sta-
ble and minimum-phase s alings D (s) to the a tual optimization formulation
is given as:

min
K (s)
min
2D
D(s)
kD(j!)Fl (P; K )(j!)D 1 (j!)k1 (8.16)

stabilizing stable, min-phase

94
In this way the optimization problem of minimizing the worst ase performan e
has been t into the H1 -synthesis framework. Optimizing over D and K si-
multaneously is in general not onvex. Therefore an indire t s heme is used in
the hope of nding a ontroller minimizing . D K-
The pro edure is alled
K (s) while holding
iteration sin e it iteratively optimizes over the stabilizing
D(s) xed and then optimizes over stable minimum-phase D(s) while holding
K (s) xed. More details on erning the pra ti al implementation of the syn-
thesis problem an be found in [18, 189. In most engineering situations the
proposed s heme has been proven to be su essful.

8.4.2 Controller synthesis and analysis

To illustrate the ee ts of the weighting fun tion sele tion, un ertainty model
sele tion, i.e the trade-o between performan e and robustness we shall study
four design ases in our example:

1. Nominal Performan e Design: All un ertainties are set to zero we desig-

nate Knom the resulting ontroller.
2. Complex Un ertainty Design: assume plant with only the omplex per-
turbation6 2
at the a tuators. The resulting ontroller will be
denoted as K2 . The augmented perturbation
related to the robust
performan e index is given by
2 = diag(6 ;
perf ).

3. Real and Complex Un ertainty Design: Taking all un ertainties into a -

ount leads to ontroller K3 . R = diag(1 ; : : : ; 5 ) 2
The real un ertainty is

R . The augmented perturbation
related to robust performan e is de-
noted as
3 = diag(1 ; : : : ; 5 ; 6 ;
perf ).

Knom K2 (Wp1 ) K2 (Wp2 ) K3 (Wp1 ) K3(Wp2 )

 (M ) 0.89 28.66 28.66 28.66 28.66
(M ) 1.71 1.86 1.88 2.11
(M ) 1.51 1.60 1.78 1.89
order 6 9 9 15 15

Table 8.2: A hieved robust performan e levels

In table 8.4.2 the a hieved robust performan e levels expressed in  values

for all ongurations are summarized. Row # 1 shows the results of the pure
H1 optimization. Rows # 2 and # 3 reveal the robust performan e levels
a hieved after the rst and se ond D K iteration. In the last row the order of
the resulting ontrollers is given. The table ree ts a well known fa t that the
robust performan e level de reases as the un ertainty and performan e levels
in rease.

95
 Controller Shape Analysis

In gure 8.10 the frequen y responses of the ontrollers are depi ted. As al-
ready noted in se tion 8.3.6, the ontroller will not ne essarily roll o at higher
frequen ies, sin e the ombination of the a tuator and the plant model al-
ready shows this behaviour. We an see that ontroller shapes atten out at
higher frequen ies and lower gains. In the se tion 8.3.6 we have seen that
the gain of K at higher frequen ies has to satisfy: jK j < 1=jW_E 15Win j =
1=(57:3=30  15  10=57:3) = 0:2. This is onrmed by gure 8.10. For Kf we
have: jKf j < 1=jW_ 15Wn j = 1=(57:3=30  15  0:1  10=57:3) = 2. This is also
E
satised.

3 T_i and Wpert

10 1
10

0
2 10 Wpert^1
10

Ti
K3 (Wp2)
1
K2 (Wp2) 10
1
10
Controller gain

K3 (Wp1)
mag

2
10
0 K2 (Wp1)
10 K_f
3
10
Knom c

1
10
K_c 4
10

2 5
10 2 1 0 1 2
10 2 1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10
frequency (rad/s) frequency (rad/s)

Figure 8.10: Controller frequen y Figure 8.11: Input ompl. sensitiv-

responses ity fun tion with
1
Wpert

Due to the a tuator un ertainty the ontroller will limit its bandwidth at
frequen ies where un ertainty starts to be ome important. This is illustrated
by the fa t that the bandwidth of the input omplementary sensitivity fun tion
is limited by the lter
1 are
Wpert . The frequen y response plots of Ti and Wpert
given in gure 8.11. Indeed Ti rolls o near the ross-over frequen y of Wpert .
In this way we prevent unmodelled higher order dynami s from ex itation by
keeping the ontrol a tions within the lower frequen y range. The ontroller
for the nominal plant in gure 8.10, Knom has only a feedforward a tion K
and no feedba k Kf = 0. The feedforward a tion K approximately inverts the
plant and the ideal model is built in as a feedforward lter. We know a priori
that in absen e of un ertainty no feedba k is required. It is interesting to see
that this out ome is a hieved automati ally by the method.

Another interesting ee t is the inuen e of the performan e weight Wp . By

in reasing the slope into the low frequen y range ( Wp1 ! Wp2 ) the ontroller
does exa tly the same. We an use this to for e the optimization algorithm
to build integration or even double integration. As in lassi al Bode loop
shape te hniques, the internal model prin iple holds and as we know in reasing
tra king error requirements requires in reasing low frequen y ontrol gain.

96
 Nominal Performan e
Beside the shape of the ontroller we are interested in the trade-os it makes.
In gure 8.12 we have for all ontrollers plotted the nominal performan e level
 (M22 ). The overall shape is, a high value of  (M22 ) at low frequen y orre-
sponding to an ee tive tra king requirement at these frequen ies. At higher
frequen ies there is no performan e requirement so that plot  (M22 ) rolls o.
Given all the ontrollers the best nominal performan e  (M22 )  0:9 is a hieved
by the system without un ertainty. For the other ontrollers the level is higher
(worse), be ause there is a trade-o against robustness to the perturbations.
We know that the omplex un ertainty is about 20 % in the low frequen y
range. This is about of the performan e degradation level of the se ond system
with respe t to the nominal system. It is surprising that the third ontroller
for the most un ertain plant a hieves a better nominal performan e level than
the se ond ontroller. The on lusion is that the real un ertainty at low fre-
quen ies, sets o the ee t of omplex un ertainty with respe t to the nominal
performan e and that this ee t is reversed at high frequen ies.
For omparison we add the nominal performan e plot for K2 designed with
Wp2 . Note that the inuen e of Wp2 is modest in the performan e level of K2 .

4
1.2
K2 (Wp2) 3.5

K2 (Wp1) Knom
1
K3 3

Knom
0.8 2.5
K2
mu(M)
sig1(M22)

2
0.6

1.5 K3

0.4
1

0.2
0.5

0 2 0 2 1 0 1 2 3
1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10
frequency (rad/s) frequency (rad/s)

Figure 8.12: Nominal performan e Figure 8.13: Robust performan e

levels ontrollers levels ontrollers

Robust Performan e
To ompare the robust performan e levels a -test for
3 is applied on all
a hieved ontrollers. The result an be found in gure 8.13. None of the
ontroller a hieves robust performan e. One of the purposes of this omparison
is to reveal the ee ts of o-design spe i ations for the ontrollers K2 and
Knom. The question we have in mind is how robust is a robustly designed
ontroller?
The nominal ontroller Knom performs worst with a 400 % degradation at
low frequen ies. At higher frequen ies it has a better robust performan e level
than K2 and K3 . The result is expe table, sin e it is purely an open loop
ontroller. The shown -plots for K2 and K3 ree t the design result in the

97
medium performan e ase, i.e. with weight Wp1 . Using Wp2 would show a too
dramati performan e ollapse of Knom ae ting the s ale of the plots. The K2
ontroller has about a 70 % performan e degradation due to real perturbations.
Note, that by taking into a ount the real perturbations in the design, ase K3 ,
the total robust performan e level improves onsiderably in the low frequen y
range at the ost of the level at higher frequen ies. There is an overall better
balan e between the performan e and robustness obje tive, whi h improves the
better we model the un ertainty in the plant.

Robust Stability
The robust stability properties are shown in gure 8.14, the values of (M11 )
are plotted along the frequen y axis. Again all perturbations, i.e.
3 , have been
taken into a ount. The ontroller Knom a hieves the best robust stability level,
whi h is not surprising anymore sin e there is no loop losure ( Kf = 0), and
apparently there is no perturbation with norm  1 to destabilize the nominal
plant (a system with no feedba k has no robust stability problems). For K2
(M11 ) < 1).
robust stability is a hieved (
Note that for K2 and K3 two bounds are visible at lower frequen ies; they
arise from real  approximations by optimizing an upper bound and a lower
bound: the exa t value of  lies in between these bounds.
If for K2 only the omplex perturbation is taken into a ount (not shown)
the plot moves approximately 0.1 downwards, whi h means that there is a 10 %
stability robustness degradation to unmodelled spe i ations in the design. It
is remarkable that the robust stability level for K3 is higher (worse) than the
one of K2 : in a small frequen y range it is even possible to nd a ombina-
tion of perturbations (  (
i )  1) that destabilizes the plant ((M11 ) > 1).
We must realize that we are optimizing the peak value of (M (j!)). Taking
the parametri un ertainties into a ount in the design improves this value
ompared to the -test for K2. In this sense we su eeded in the third de-
sign. However, the balan e between performan e and stability robustness has
moved in the wrong dire tion: the rst improved, the se ond got worse, while
overall robust performan e improved. The designer has to be areful and has
always to nd a right balan e. Espe ially, sin e in pra ti e (M ) < 1 (8 !)
is seldomly a hieved. However, for a given situation robust stability should be
preferably the rst to be guaranteed, i.e. ( (M11 (j!)) < 1 8 !). Then, slowly
and arefully, the designer an buy performan e from the robust stability until
the overall requirements are in balan e. We would like to remark that in the
multivariable ase this philosophy still holds but things be ome more omplex
be ause of dire tionality issues.

Time Simulations, Performan e Sensitivity

We shall next analyze the systems performan e via pit h rate step ommand
simulations shown in gure 8.17. Three model ases are onsidered:

 nom: simulation with the nominal model;

98
 1.2

0.8 K3

K2
mu(M11)
0.6

0.4

0.2 Knom

0 2 1 0 1 2 3
10 10 10 10 10 10
frequency (rad/s)

Figure 8.14: Robust stability ontrollers

 pert1: simulation with a perturbed model: all parameters in table 8.1 are
perturbed with = 20%; only ME with = +20%;
 pert2: identi al perturbation, now
ME = 20% the others are 20%.
The rst plot shows the rst ontroller: the nominal response oin ides well
with the ideal model response. The perturbations have a dramati ee t on the
tra king performan e, sin e we are in fa t looking at an open loop simulation.
In the responses for the other ontrollers the ee ts are less dramati . Note
that for the se ond ontroller K2 designed with Wp2 the steady state error
indeed has be ome nearly zero, even under the inuen e of the perturbations.
Finally, we an see that K3 performs better than K2 (with Wp1 ) under the
perturbations, as ould be expe ted from the -analysis.

Time Simulations, Robust Stability Aspe t

Finally, we are looking at the perturbations that ould destabilize the losed
loop systems in the ase of K2 and K3 . We an see in gure 8.14 that the  plots
(lower bounds) of the losed loop systems show peak values of peak = 0:918 at
!peak = 6:5 rad/s and peak = 1:024 at !peak = 6:9 rad/s respe tively. This
means that we an nd the smallest destabilizing perturbation with appropriate

-stru ture:

= fdiag[1;


; 6 : 1 ;


; 5 2 IR ; 6 2 Cg I

with  (
) = 1=0:918 = 1:089 and  (
) = 1=1:024 = 0:977 respe tively. For
the robust stability test we have a norm bounded
in mind,  (
)  1
2
.
The system with K2 is robustly stable, sin e peak < 1 and the norm of the
perturbation therefore needs to be larger than 1 to destabilize the system. This

99
is not the ase for K3 . Using available software we have found perturbations
that will just destabilize the systems. For K2 we have (for example):

= diag[1;


; 6
= diag[ 1:0892; 0:8389; 0:7893; 1:0892; 1:0892; 1:0573 0:2618i
with norm 
 (
) = 1:098. For K3 :

= diag[ 0:9768; 0:9768; 0:0073; 0:9768; 0:9768; 0:9624 0:1667i
In gure 8.15,8.16 we simulate the losed loop system, without and with per-
turbations. To see how sensitive results an be, we also implement the pertur-
bation s aled to 98 % and 1:02 % of its riti al value. We an see that both
systems with ontrollers K2 and K3 indeed are destabilized while in reasing
the perturbation levels over their riti al values, whi h on ludes the example.

controller K2
0.35 controller K3
0.3

0.3 nominal
0.98*pert 0.25
1.02*pert
0.25 1*pert
0.2

0.2
q (deg/s)
q (deg/s)

0.15

0.15

0.1

0.1
nominal
0.98*pert
0.05
1.02*pert
0.05
1*pert

0
0 0 1 2 3 4 5 6 7 8 9 10
0 1 2 3 4 5 6 7 8 9 10 time (t)
time (t)

Figure 8.15: Destabilizing pertur- Figure 8.16: Destabilizing pertur-

bations bations

Although this example is very simple, it is lear that -synthesis is a pow-

erful tool where many fa tors an be taken into a ount: design requirements,
un ertainties, disturban e models et . In absen e of un ertainty, the two de-
grees of freedom ontroller inverts the plant and pla es in the feedforward path
almost no feedba k. This is a desirable strategy only in the absen e of un-
ertainties. We saw that slight un ertainties aused huge performan e degra-
dation. The di ulty in designing a good ontroller is the to nd the right
trade-os between the many usually oni ting requirements. We believe that
the approa h as shown here provides us with tools to make sensible (balan ed)
design de isions to a hieve robust performan e. We on lude by saying that
the method for es the designer to understand his model and the intimately re-
lated spe i ations on it. The method links the design work with the pra ti al
world.

8.5 Con lusion

We have reviewed a general framework for ontrol system analysis and syn-
thesis. The stru tured singular value  arose from the stability analysis of a

100
more general type of systems, namely linear fra tional transformations. This
permits us to ta kle formally the robust performan e paradigm. To over ome
the often reported di ulties in the la k of guidan e in the weighting fun tion
sele tion during the design we have provided a simple and illustrative example
whi h ontains all ingredients and steps that should be arried out in analyz-
ing su h a design problem. Hopefully, we have su eeded in larifying that a
good design is a matter of balan ing requirements. We might say that  is
the tool to guide us in nding the required trade-os between performan e and
robustness. It pla es the hallenge on the side of the pra ti ing engineer. To
be su essful in improving the behaviour of omplex systems he will have to
quantify his spe i ations and he will have to rely ontinuously on a better
and deeper system knowledge. The paradigm is no longer ontroller design,
but spe i ation design.

Knom K2 , with Wp1

14 12

pert1
12 pert1
10
nom
ideal
10 pert2
nom 8
ideal
8
q (deg/s)

q (deg/s)

pert2 6

4
4

2
2

0 0
0 5 10 15 0 5 10 15
time (s) time (s)

K2 , with Wp2 K3 , with Wp1

12 12

nom, pert1
10 10
pert 1
pert2
8 8
pert2
q (deg/s)

q (deg/s)

6 6

4 4

2 2

0 0
0 5 10 15 0 5 10 15
time (s) time (s)

Figure 8.17: Step responses for designed ontrollers

101
9. Nonlinear Dynami Inversion Control

Binh Dang Vu 1

9.1 Introdu tion

Among the spe i methodologies for the ontrol of systems des ribed by non-
linear mathemati al models, dynami inversion is ertainly the most widely
investigated by ontrol resear hers in the last two de ades. A omplete theory
is now available for the design of feedba k ontrol laws whi h render ertain
outputs independent of ertain inputs (disturban e de oupling and noninter-
a ting ontrol) or whi h transform a nonlinear system into an equivalent linear
system (feedba k linearization or dynami inversion).
The theory of nonintera ting ontrol was initiated by the pioneering work
on linear systems by Falb and Wolovi h [75. The extension to nonlinear sys-
tems is due to the work of Singh and Rugh [212, Freund [83, following an
idea due originally to Porter [196. Feedba k linearization is based on some
early work of Krener [139 and Bro kett [34 demonstrating that a large lass
of nonlinear systems an be exa tly linearized by a ombination of a nonlinear
transformation of state oordinates and a nonlinear state feedba k ontrol law.
A major breakthrough o urred at the beginning of the eighties with the appli-
ation of mathemati al on epts derived from the eld of dierential geometry,
Isidori et al. [125, Byrnes and Isidori [42. A good survey of the theory an
be found in re ent books : Isidori [124, Nijmeijer and Van Der S haft [184,
Slotine and Li [216.
The basi feature of feedba k linearization is the transformation of the orig-
inal nonlinear ontrol system into a linear and ontrollable system via a non-
linear state spa e hange of oordinates and a nonlinear stati state feedba k
ontrol law. The solution of this problem relies on the nonsingularity of the
so- alled de oupling matrix. When this ondition is not satised, a dynami
state feedba k ontrol law an be investigated. Su ient onditions for dy-
nami feedba k linearization have been given by Fliess [80 who introdu ed
the dierential rank of a system. The dierential rank measures the degree of
independen e of the system outputs and an be onsidered as the nonlinear
equivalent of the rank of the transfer matrix for a linear system.
When the ondition of nonsingularity is satised by the given system (stati
feedba k) or by a suitable extension of the given system (dynami feedba k),
the feedba k ontrol law an be omputed by solving a set of state independent

1
O e National d'Etudes et de Re her hes Aerospatiales (ONERA), BA701, 13661 Salon
de Proven e Air

102
algebrai linear equations. This is a result of the stru ture of the dynami s
whi h is assumed to be ane in the ontrols.
As the input-output behaviour of the resulting state-feedba k system resem-
bles that of a linear time-invariant system, any linear ontrol design te hnique
an be applied to a hieve the design performan e. However, in order to guaran-
tee the internal stability of the system, it is not su ient to look at input-output
stability, sin e all internal unobservable modes of the system must be stable
as well. The rst step in analysing the internal stability of the system is to
look at the zero dynami s. The zero dynami s of a nonlinear system are the
internal dynami s of the system subje t to the onstraint that the output, and
therefore all the derivatives of the output, are set to zero for all time.
There have been many appli ations of nonintera ting ontrol and feedba k
linearization to air raft ight ontrol problems : Asseo [15, Singh and S hy
[213, Meyer and Ci olani [170, Dang Vu and Mer ier [51, Menon et al. [168,
Lane and Stengel [149, Bugajski et al. [41, Adams et al. [6.... The main
advantage of the feedba k linearization te hnique is that it does not require gain
s heduling to ensure ight ontrol system stability over the entire operational
envelope of the air raft. Traditional air raft ontrol designs have to rely on
linearized models obtained throughout the ight envelope of the vehi le, with
linear ontrollers synthesized for the set of resulting linearized models.

9.2 Plant Model Requirements and Controller

Stru ture
9.2.1 SISO ase
The essentials of the approa h are most easily understood in terms of the simple
single-input single-output problem.
The method of synthesis onsiders a lass of nonlinear systems ane in
ontrol

x_ = f (x) + g(x)u (9.1)

y = h(x) (9.2)

where f , g are smooth ve tor elds on IR n and h is a smooth fun tion mapping
IR
n ! IR .
Su h a system is feedba k linearizable of relative degree r if there exist state
and input transformations

z = (x) z 2 IR r (9.3)

u = (x) +  (x)v v 2 IR (9.4)

where  (x) 6= 0 and  is a dieomorphism whi h transforms (9.1) into a on-

trollable linear system
z_ = Az + Bv (9.5)

103
 Indeed, we time-dierentiate (9.2) to obtain

h
y_ = (f (x) + g(x)u) (9.6)
x
If the oe ient of u is zero, we dierentiate (9.6) and ontinue in this way
until a nonzero oe ient appears. This pro ess an be su in tly des ribed
by introdu ing some onventional notation of dierential geometry. The Lie
derivative of the s alar fun tion h with respe t to the ve tor eld f is dened
as
h
Lf h(x) = f (x) (9.7)
x
Higher order derivatives may be su essively dened

Lkf h(x) = Lf (Lfk 1 h(x)) (9.8)

With this notation, (9.6) an be written

y_ = Lf h(x) + Lg h(x)u (9.9)

If Lg h(x) = 0, then dierentiate (9.9) to obtain

y = L2 h(x) + Lg Lf h(x)u
f (9.10)

If Lg Lfk 1 h(x) = 0 for k = 1; :::; r 1, but Lg Lrf 1h(x) 6= 0, then the pro ess
ends with
y(r) = Lrf h(x) + Lg Lrf 1h(x)u (9.11)

The number r is alled the relative degree of (9.1).

Now if we dene the oordinates z 2 IR r
zk = k (x) = Lk 1 h(x) k = 1; :::; r
f (9.12)

then we get the linear r-dimensional ompletely ontrollable and observable,

ompanion form system

0 1 0 1
0 1 0 : 0 0
B0
B 0 1 0 :CC
B0C
B C
z_ = B
B : : 0 1 0CC z + B : C v = Az + Bv
B C (9.13)
 : : : : 1 A 0A
0 0 : : 0 1
where
v = Lrf h(x) + Lg Lrf 1 h(x)u (9.14)

Su h a system is alled a Brunovsky anoni al form. Exa t linearization is

a hieved when the relative degree is equal to the system order (r = n).
The ontrol law is obtained by transforming the above linear system state
variables and ontrol into the original oordinates, with ontrol

u = (x) +  (x)v (9.15)

104
where
Lrf h(x) 1
(x) =  (x) =
Lg Lrf 1h(x) Lg Lrf 1 h(x)
(9.16)

The ontrol law v is hosen depending on the ontrol task. For instan e, if
y is required to be stabilized around zero, we hoose v as
r 1
X
v= k zk+1 (9.17)
k=0
in order to a hieve the design performan e for the output dynami whi h is
given by
y(r) + r 1 y(r 1) + ::: + 1 y(1) + 0 y = 0 (9.18)

Stabilization of (9.18) annot guarantee stabilization of (9.1). A omplete

hara terization of the stability properties of (9.1) requires a view of the entire
state spa e. The key result of Isidori [124 is that there exists a transformation
of oordinates whi h provides a so- alled normal form for (9.1), from whi h a
omplete stability pi ture an be obtained

x ! (z; ) z 2 IR r  2 IR n r (9.19)

z_ = Az + Bv (9.20)

_ = q(z; ) (9.21)

The zero dynami s of the system (9.1) are dened by the equation

_ = q(0; ) (9.22)

whi h orresponds to the internal behaviour of the system when the ontrol is
hosen to onstraint the output to be identi ally null.
For tra king ontrol problems, for instan e if y is required to tra k yd, we
hoose v as
r 1
X
v = yd(r) k (zk+1 yd(k) ) (9.23)
k=0
in order to a hieve the design performan e for the tra king error

e = y yd (9.24)

whose dynami is given by

e(r) + r 1 e(r 1) + ::: + 1 e(1) + 0 e = 0 (9.25)

Again the internal behaviour must be bounded. It an be shown that for

any  > 0, there exists so that
je(k) (t0 )j < k = 0; :::; r 1 =) je(k) (t)j <  8t > t0 > 0 (9.26)

k(t0 ) R (t0 )k < =) k(t) R (t)k <  8t > t0 > 0 (9.27)

where _R = q(zR ; R ) and zR = (yd ; yd(1) ; :::; yd(r 1))T

105
9.2.2 MIMO ase
The multi-input multi-output ase is qualitatively similar to the single-input
single-output ase.
Consider a nonlinear dynami al system in the form

x_ = f (x) + g(x)u (9.28)

y = h(x) (9.29)

where x 2 IR n , u 2 IR m , y 2 IR m , and f , g and h are smooth fun tions of x.

The problem onsists of nding m transformations of oordinates and a ontrol
law

z i = i (x) z i 2 IR ri i = 1; :::; m (9.30)

u = (x) +  (x)v v 2 IR m (9.31)

where ri is the relative degree asso iated to the output yi , whi h transform
(9.28) into an equivalent ontrollable linear system

z_i = Ai z i + Bi v i = 1; :::; m (9.32)

from whi h the auxiliary ontrol synthesis is performed.

Under the ondition of nonsingularity of the matrix

ij (x) = Lgj Lfri 1 hi (x) i = 1; :::; m j = 1; :::; m (9.33)

the linearizing oordinates are given by

zki = Lfri 1 hi (x) i = 1; :::; m k = 1; :::; ri (9.34)

and the ontrol law u is obtained from

(x) =
1 b  (x) =
1 (9.35)

with
bi = Lrfi hi (x) i = 1; :::; m (9.36)

The ontrol law v is hosen depending on the ontrol task. For instan e, if

rXi 1
(r
vi = ydi i)
ik (zki +1 yd(ki ) ) (9.37)
k=0
then we obtain a nonintera ting ontrol system whi h performs a de oupled
tra king of yd by y, omponent by omponent. In this ase, the matrix
is
alled the de oupling matrix.
The input-output behaviour is dened by the diagonal transfer matrix

1
H (s) = diag( ) i = 1; :::; m (9.38)
di (s)

106
with
di (s) = i0 + i1 s + ::: + iri 1 sri 1 + sri (9.39)

The stru ture of a simple ontrol system (ri = 1 i = 1; :::; m) is depi ted
in Figure 9.1. As the output y is required to tra k the ommanded value yd ,
we hoose v as
v = !(y yd) (9.40)

where
! = diag( i0 ) i = 1; :::; m (9.41)

The ontrol law is then given by

h h h
u= ( g(x)) 1 !(y yd ) ( g(x)) 1 f (x) (9.42)
x x x

h -1 h
( g) f
x x

yd + v h -1 u . x
( g) x=f(x)+g(x)u
x +

y h(x)

=diag(c0i) i=1,...,m

Figure 9.1: Controller stru ture

9.3 Possible Design Obje tives and Design Cy le

Des ription
A straightforward appli ation of the linearization te hniques might result in a
system :

 with unstable unobservable modes resulting in undesirable losed-loop

system behaviour;

 with large ontrol eort leading to ontrol saturation.

Preliminary physi al onsiderations are ne essary to obtain a good design.

By negle ting ertain physi ally small variables, the approximate linearization
might result in better performan e. Singular perturbation theory an also be

107
used to reformulate the original dynami model as two or more lower-order sys-
tems whi h are better onditioned for linearization; a ommon example is the
time-s ale separation between the translation dynami s of an air raft and its
rotational dynami s. Good zero-order dynami behaviour and redu ed ontrol
a tivity rely on a good hoi e of the ontrolled variables and their dynami s
(ve tor v in the design).
The overall approa h for the ontrol design is as follows :

 Step 1. Reformulate if ne essary the original dynami system to obtain

an approximate nonlinear model for whi h a state-dependent nonlinear
inverse an be easily onstru ted.

 Step 2. Constru t the nonlinear inverse. The losed-loop system formed

by the nonlinear inverse and the approximate nonlinear model redu es to
a linear time invariant system.

 Step 3. Use any suitable linear ontrol design te hnique to synthesize a

ontroller for the above linear system (e.g. eigenstru ture assignment).

 Step 4. Transform the linear system state variables and linear ontrol
into the original oordinates and ontrol.

 Step 5. Iterate on linear dynami hara teristi s to obtain required per-

forman e with redu ed ontrol a tivity.

 Step 6. Eventually omplete the design by synthesizing a robust ontroller

using adequate linear te hniques.

9.4 A Simple Design Example

The following example on erns the ontrol problem of a simplied V/STOL
air raft model and is taken from Meyer et al. [171. A simple air raft is used
that has a minimum number of states and inputs, but retains many of the
features that are onsidered when designing ontrol laws for a real air raft
su h as the Harrier. Figure 9.2 shows the PVTOL (planar verti al takeo and
landing) air raft, whi h is the natural restri tion of V/STOL air raft to jet-
borne operation (e.g. hover) in a verti al plane. The air raft state is simply the
position y, z , of the air raft entre of mass, the angle  of the air raft relative
to the y-axis, and the orresponding velo ities, y_ , z_ , _ . The ontrol inputs, u1 ,
u2 , are the thrust (positive downward) and the rolling moment.
The equations of motion for the PVTOL air raft are given by

y = u1 sin  + u2 os  (9.43)

z = u1 os  + u2 sin  1 (9.44)

 = u2 (9.45)

where -1 is the gravitational a eleration and  is a small oe ient giving
the oupling between the rolling moment and the lateral a eleration of the
air raft.

108


Figure 9.2: The planar verti al takeo and landing air raft

Choosing y and z as the outputs to be ontrolled, we seek a state feedba k

law of the form

u = (x) +  (x)v (9.46)

su h that, for some r = (r1 ; r2 )T ,

y(r1) = v1 (9.47)

z (r2) = v2 (9.48)

Here, v is our new input and x is used to denote the entire state of the system.
Pro eeding in the usual way, we dierentiate ea h output until at least one
of the inputs appears. This o urs after dierentiating twi e and is given by

      
y 0 sin   os  u1
= + (9.49)
z 1 os   sin  u2
Sin e the de oupling matrix is nonsingular (although almost singular as
its determinant is ), we an linearize the system by hoosing the stati state
feedba k law

       
u1 =
sin  os  0
+
v1
u2 os sin  1 v2
(9.50)
 

109
 The resulting system is

y = v1 (9.51)

z = v2 (9.52)
1
 = (sin  + v1 os  + v2 sin ) (9.53)

This feedba k law makes the input-output map linear, but has the unfortunate
side-ee t of making the dynami s of  unobservable. Constraining the outputs
and derivatives to zero by setting v1 = v2 = 0, the zero dynami s are found to
be
1
 = sin  (9.54)

Equation (9.54) is simply the equation of an undamped pendulum. Nonlin-
ear systems, su h as (9.51)-(9.53), with zero dynami s that are not asymptoti-
ally stable are alled non-minimum phase.
From the above analysis, it is lear that exa t input-output linearization
of a system an lead to undesirable results. The sour e of the problem lies in
trying to ontrol modes of the system using inputs that are weakly ( ) oupled 
rather than ontrolling the system in the way it was designed to be ontrolled
and a epting a performan e penalty for the parasiti ( ) ee ts. For the sim-
ple PVTOL air raft, the linear a eleration should be ontrolled by ve toring
the thrust ve tor (using moments to ontrol this ve toring) and adjusting its
magnitude using the throttle. The PVTOL air raft is now modelled as

ym = u1 sin  (9.55)

zm = u1 os  1 (9.56)

 = u2 (9.57)

so that there is no oupling between rolling moments and lateral a eleration.

Dierentiating the model system outputs, ym and zm , we get
      
ym 0 sin  0 u1
= + (9.58)
zm 1 os  0 u2
Now, however, the de oupling matrix is singular whi h implies there is no
u2 enters the system
stati state feedba k that will linearize (9.55)-(9.57). Sin e
through  , we must dierentiate (9.55)-(9.56) at least two more times
      
ym (4) = u1_ 2 sin  2u_ 1_ os  + sin  u1 os  u1
zm (4) u1 _ 2 os  2u_ 1_ sin  os  u1 sin  u2
(9.59)
The de oupling matrix is invertible as long as the thrust, u1 is nonzero.
Physi ally, this result in agreement with the fa t that no amount of rolling will
ae t the motion of the air raft if there is no thrust to ee t an a eleration.
Linearizing the above system using the dynami state feedba k law
!
u1 _ 2
    
u1 sin  os  v1
= 2u_ 1 _ + os sin  (9.60)
u2 u1 u1 u1 v2

110
results in

ym(4) = v1 (9.61)

zm(4) = v2 (9.62)

Unlike the previous ase, the linearized model does not ontain any unob-
servable zero dynami s. Thus, using a stable tra king law for v, we an tra k
an arbitrary traje tory and guarantee that the model will be stable.
Of ourse, the natural question that omes to mind is : will a ontrol law
based on the model work well when applied to the true system? If  is small
enough, then the system will have reasonable properties, su h as stability and
bounded tra king.
This example shows that preliminary physi al onsiderations are ne essary
to obtain a good design. By negle ting ertain variables whi h are physi ally
small, the approximate linearization results in better performan e.

9.5 Con lusion

Feedba k linearization or dynami inversion has drawn onsiderable attention
over the last two de ades and oers a potentially powerful alternative ontrol
design methodology. Dynami inversion is an attra tive te hnique as it avoids
gain s hedules. Instead, it uses dynami models and full-state feedba k to
globally linearize dynami s of sele ted ontrolled variables. Simple ontrollers
an then be designed to regulate these variables with desirable losed loop
dynami s.
Theory of feedba k linearization is still gradually developing. There are lim-
itations and open problems. The main drawba k might be that modes be ome
unobservable under the linearization or de oupling onstraints, whi h an be
unsurmountable in ase they are unstable. The dimension of the unobservable
manifold and the omplexity of the ontrol law an vary drasti ally a ording to
the assumptions made on the model used (e.g. small oupling terms negle ted
or not). Preliminary physi al onsiderations are then ne essary to obtain a
good design. The design method requires, more or less, a urate knowledge of
the state of the system, while no satisfa tory theory for the design of the non-
linear observers is available. A suitable nonlinear analogue of the separation
prin iple still needs to be developed. One area of resear h, already initiated,
is that of ombining the design te hnique developed so far, with appropri-
ate robust te hniques whi h ould take into a ount unknown parameters and
unmodelled dynami s : LQ, -synthesis, QFT, Lyapunov synthesis, adaptive
ontrol, dierential games.

111
10. Robust Inverse Dynami s Estimation

Ewan Muir 1

10.1 Introdu tion

Robust Inverse Dynami s Estimation (RIDE) [48, [33,[176 has developed
from two other methods: the Salford Singular Perturbation Method [115 and
Pseudo-Derivative Feedba k [194, [14. Both of these methods use the same
type of multivariable Proportional plus Integral (PI) ontroller stru ture but
use a high gain to provide the desired de oupling and losed-loop dynami s.
RIDE is a development of both these methods whi h repla es the high gain
with an estimate of the inverse dynami s of the air raft with respe t to the
ontrolled outputs. This inverse input gives RIDE strong similarities to Non-
linear Dynami Inversion [218 and is similar to the equivalent ontrol found in
Variable Stru ture Control [244.

10.2 General Stru ture

The RIDE ontroller onsists of 3 omponents: a model inverse input, a PI
ontroller and a feedforward, as shown in Fig. 10.1 below.

v
_
KV Kudi
T-1 .
+
v ^udi
x
yc
+ . +
y
+ KI r r u
+ + + Aircraft
_ _

KP

Figure 10.1: Stru ture of RIDE ontrol law

1
Defen e Resear h Agen y, Flight Dynami s and Simulation Department, Bedford, MK41
6AE, UK

112
 - The model inverse provided by the dynami inverse input, u^di , a ts to
de ouple the outputs from ea h other and from the other air raft states by using
moment an ellation. The inverse is uniquely for the outputs to be ontrolled
and is therefore for a subset of the omplete air raft model only.
- Having inverted the air raft model with respe t to the outputs and de ou-
pled these, the PI ontroller then assigns to the outputs, the dynami s desired
by the ontrol law designer. The integral a tion, with gain KI , provides robust-
ness against errors in the estimate of the model inverse. The proportional gain
matrix KP provides stability and is positioned su h that it provides pseudo-
derivative feedba k.
- The feedforward omponent, onsisting of a washout lter on the demands
and des ribed by equation 10.1, is used to tune the step response hara teristi s
to give an appropriate onset of response.

feedforward input = [(T s + I ) 1 KV sy (10.1)

where T is a diagonal matrix of washout time onstants and KV is a matrix of

gains on the feedforward inputs.
The stru ture provides de oupling between the outputs and assigns them a
se ond order response whi h is spe ied by the designer. The transient response
shape an be adjusted using the feedforward.
The design method provides simple me hanisms for handling motivator po-
sition and rate limiting and it is anti ipated that motivator redundan y an be
handled separately from the ontroller design.

10.3 Closed-loop System Chara teristi s and

Gain Cal ulation (Output Feedba k Case)
The poles of the losed-loop system with a RIDE ontrol law are determined
by the following:
- the open loop system transmission zeros,
- the feedforward washout time onstants given in matrix T;
- the eigenvalues of the matrix [s
2I + CBKP s + CBKI .
As the rst set of poles oin ides with the transmission zeros of the open loop
system, it is ne essary to ensure that the feedba k measurements sele ted give
stable transmission zero lo ations. The se ond set is spe ied by the designer
who sele ts the feedforward time onstants. The third set of eigenvalues an be
assigned to the poles of a standard se ond order transfer fun tion of the form
given in equation 10.2, through appropriate gain sele tion.

y = (s2 I + 2Zd
n s +
2n ) 1
2n y (10.2)

The proportional, integral and feedforward gain matri es, KP , KI and KV

respe tively, are al ulated from the inverse of the motivator ee tiveness ma-
CB ) 1 , the matrix of desired losed system damping, Zd, natural fre-
trix, (
quen y,
n , and feedforward gain, M , where Zd ,
n and M are diagonal ma-
tri es spe ied by the ontrol law designer.

113
 KP = (CB ) 1 2Zd
n (10.3)

KI = (CB ) 1
2n (10.4)

KV = (CB ) 1 M (10.5)

For the output feedba k ase and using the gains al ulated in equations
10.3 to 10.5 above, the output equation for the losed-loop system is des ribed
by equation 10.6.

y = (s2 I +2Zd
n s +
2n) 1 [
2n y + sM (T s + I ) 1y + s(CB )(^udi udi) (10.6)

As Zd,
n , T and M are all diagonal matri es, ea h demand y will ae t
only one of the outputs y. Therefore, for the output feedba k ase, RIDE will
provide tra king of the demands with unity steady-state gain, the dynami s
of the response an be spe ied and will be se ond order. The feedforward
omponent, whose dynami s are spe ied by the matri es T and M , will shape
the initial response to any inputs. Any errors in the estimate of the dynami
inverse input will be orre ted by the integral loop. The rate at whi h this
happens will be dependent on the integral gain KI .
The role of the dynami inverse input, udi , is to keep y_ = 0 and thus an
estimate, u^di , is al ulated from
u^di = (CB ) 1 CAx (10.7)

where the matrix ( (CB )

1 CA) is represented by the gain matrix Kudi in Fig.
1. Note that the state ve tor x need only ontain the rigid body states whi h
dire tly ae t the outputs y .

10.4 Design Limitations

RIDE does not take into a ount expli itly any motivator or sensor dynami s
during the design phase. In many ases, the dynami s of the a tuators and
sensors will be su iently fast to maintain the desired performan e. Should
this not be the ase, a areful hoi e of the design parameters will re tify the
problem. For example, the spe ied bandwidth of the losed-loop response
ould be redu ed and the feedforward used to maintain an adequate speed of
response.

10.5 Controller Synthesis Aspe ts

The simpli ity of RIDE naturally means that it does not provide the ompre-
hensive solution promised by other more omplex methods. RIDE does not

114
provide expli it guarantees in terms of either stability or performan e robust-
ness. It is also limited in terms of the amount of spe i ation data whi h an
be in orporated dire tly in the design stage. Therefore separate analysis is
required on e the initial design has been done, to see if the ontroller meets
the spe i ation. However, in pra ti e, RIDE has been found to produ e on-
trollers with a eptable time responses, even when performing highly dynami
manoeuvres with non-linear air raft models [176, and it is possible for design-
ers to a hieve satisfa tory gain and phase margins. Also the integral a tion
provides robustness to errors in the dynami inverse input estimate.
The simpli ity of RIDE, both in terms of the underlying mathemati s and
the design pro ess, means that the learning urve is short and undemanding.
Also, no spe ialist skills, design software or omputer hardware are required
and the resulting ontroller is simple with a lear stru ture.
A fuller understanding of the ontroller synthesis pro edure and of the de-
sign method an be obtained from the RIDE-HIRM ontrol law in hapter
33.

115
11. A Model Following Control
Approa h

Holger Duda1 , Gerhard Bouwer1 , J.-Mi hael Baus hat1

and Klaus-Uwe Hahn 1

11.1 Introdu tion

Design and development aspe ts for state of the art ontrol systems are based
today on improved system models and omputer supported tools. One of the
design aims for a ontrol system is a low feedba k authority. High feedba k
gains, espe ially in multi input/multi output systems with un ertainties, may
lead to stability problems, whi h are often di ult to predi t. A well-known
re ipe to avoid this is:

Put all available information about the pro ess to be ontrolled into
the feedforward bran h of your ontrol system.

In view of ight ontrol system design it an be assumed that a detailed

database of the air raft to be ontrolled is available, be ause it is usual to im-
prove aerodynami databases of new air raft during ight testing using modern
system identi ation te hniques [109. Therefore, it is highly re ommended to
use this knowledge in the ight ontrol system design pro ess [108.
One of the most promising approa hes, whi h takes the above mentioned
aspe ts into a ount, is the Model Following Control (MFC) te hnique. Even
in the early stages of ight ontrol system resear h MFC on epts seemed to be
promising [243. An improved theoreti al understanding of the identi ation
of system dynami s promotes the appli ation of MFC systems [107.
The design problem for the MFC on ept an be separated into three inde-
pendent subtasks: First a ommand blo k has to be dened, whi h in ludes the
desired dynami behaviour taking into a ount the plant's performan e limits.
Then a ontroller omplex onsisting of a feedforward and a feedba k ontroller
has to be designed, whi h is independent of any ommand blo k hara teristi s,
su h as manual ight ontrol laws or autopilot fun tions.
The expression  ommand blo k has been introdu ed instead of model in
order to avoid misunderstandings on erning the plant model. It predominantly
denes the dynami behaviour of the overall MFC system. The feedforward
ontroller ontains an inverse model of the plant to be ontrolled. Assuming no
external disturban es, a perfe t knowledge and an ideal inversion of the plant,
1
DLR German Aerospa e Resear h Establishment, Institute of Flight Me hani s, D-38108
Brauns hweig

116
the omplete ontrol ould be performed by the feedforward ontroller without
any feedba k a tivity. In pra ti e, a feedba k ontroller is required, whi h
only has to manage the remaining ontrol part not overed by the feedforward
bran h, whi h will always perform the majority of ontrol a tivity.
Sin e the aim of the RCAM Design Challenge is to evaluate ontrol theories
on erning robust ight ontrol system design, one has to answer the rising
question:

How an the MFC on ept ontaining high authority feedforward

ontrol enhan e robustness, whi h is mainly ae ted by feedba k
ontrol?

The feedforward part of the MFC represents a kind of partly inherent robust-
ness ompared with a pure feedba k system. By an exa t denition of the
desired performan e and the limitations of the pro ess one omes to oni t-
free ontrol a tions and, therefore, to minimum feedba k ontrol a tivity for
manoeuvres. This leaves maximum authority to the feedba k ontroller to ope
with un ertainties and disturban es reje tion.
Besides these robustness aspe ts the MFC on ept provides several addi-
tional benets regarding pra ti al appli ations:

 The ommand blo k denes predominantly the input ommand behaviour

of the overall MFC system. Therefore, tailored Flying Qualities or au-
topilot fun tions an be easily realised.

 The feedforward and feedba k ontrollers are independent from the lay-
out of the ommand blo k. This separation allows a lear sharing of
responsibilities for the design tasks with well dened interfa es, whi h
an be performed by dierent teams. Therefore, design problems whi h
may be observed during simulator or ight testing an be easily lo ated
and solved.

 The overall ontroller stru ture allows the denition of ertain ommand
blo k modules for spe ial tasks, su h as manual ight ontrol laws or
autopilot fun tions for an air raft family. A re-design for other (similar)
air raft does not have to go through all the individual steps, but only the
feedforward and feedba k ontrollers have to be adapted. The attainable
ommonality of ying hara teristi s for an entire ategory of air raft
type is a protable element onsidering pilot training and erti ation
aspe ts.

11.2 Typi al Appli ations

MFC on epts have been utilised in ight systems sin e the sixties [243. The
main appli ation of the MFC approa h is in-ight simulation. The aim of
in-ight simulation is to impose the hara teristi s of a ight vehi le to be sim-
ulated on airborne simulators, su h as Calspan's TIFS (Total In-Flight Simu-
lator) [175, DLR's ATTAS (Advan ed Te hnologies Testing Air raft System)

117
and ATTHeS (Advan ed Te hnologies Testing Heli opter System) [40. Fur-
thermore, MFC on epts have been realised in several experimental heli opter
programs in the United States and were even hosen for the new operational
Fly-by-Wire heli opter Coman he [93.
The appli ation potential of MFC systems is demonstrated below by re ent
resear h programs arried out at DLR Institute of Flight Me hani s. They
have been hosen be ause all have been ight-demonstrated in a real-time and
real-world environment.

11.2.1 In-ight simulation

Various in-ight simulations have been arried out in the xed-wing and heli-
opter area, su h as the Airbus A3XX air raft, the Indonesian N250 air raft,
the HERMES Spa eplane, and the Lynx heli opter.
One of the most re ent appli ations of ATTAS has been the airborne sim-
ulation of an Airbus A3XX-type transport air raft. The airframe model is
based on preliminary data of the unaugmented air raft without elasti modes.
A typi al ight-test result is illustrated in gure 11.1. Manually own turn
reversals learly show the ex itation of the low damped dut h roll of the im-
plemented model. The time histories demonstrate a good mat hing between
the A3XX model states (solid lines) and the measured ATTAS states (dashed
lines) validating the model following me hanism.

Roll Command of the Pilot (deg)

10

-10
Bank Angle (deg)
20
0

-30
Angle of Sideslip (deg)
5

0
A3XX
ATTAS
-6
0 100 200 300 Time (sec) 400

Figure 11.1: A3XX In-Flight Simulation (Flight-Test Results)

Espe ially in the heli opter area the MFC on ept has been proven to be a
very valuable tool due to the highly ompli ated ouplings of basi heli opter
dynami s [104. The in-ight simulation of the Lynx heli opter shall serve as
an example [31. This heli opter has ouplings opposite to the orresponding
ouplings of ATTHeS in its basi BO 105 mode. Flight tests have been arried
out demonstrating, that all ATTHeS states mat h well the ommanded Lynx

118
model states. In general, the in-ight simulation was deemed by the pilots to
be representative for the Lynx heli opter.

11.2.2 Flight ontrol system resear h

Flight ontrol system resear h proje ts based on MFC on epts have been
performed, whi h are summarised below:

 LADICO (Lateral/Dire tional Control of an Air raft):

This proje t was arried out within the framework of the GARTEUR A -
tion Group FM (AG 06) Low-Speed Lateral/Dire tional Handling Quali-
ties Design Guidelines. A lateral/dire tional ontrol system for a trans-
port air raft was developed, featuring an MFC on ept [38. Piloted
evaluations of the system in two ground based simulators demonstrated
its performan e; all evaluation pilots rated the system as Level 1.

 ARCORE (Arti ial Redundan y Con ept for Re onguration):

Flight ontrol system re onguration on epts have been developed and
ight tested [21. The investigated failure was an elevator stu k in the
trim position, whi h has been ompensated by the re onguration on-
troller using the stabiliser with its poor dynami s for altitude ontrol
instead of the stu k elevator.

 SCARLET (Saturated Command and Rate Limited Elevator time delay):

Air raft-Pilot Coupling (A-PC) problems due to rate saturation have
been investigated and ight tested. An alternative ontrol s heme has
been developed, whi h ompensates for the additional time delay due to
rate saturation. The ight test results were very promising [39.

 ADS-33D riteria (Aeronauti al Design Standard):

Flying Qualities Databases for modern Fly-by-Wire heli opters have been
developed on ATTHeS [31.

11.3 Plant Model Requirements

The MFC approa h is well tuned for the usually available pro ess knowledge
in the ight ontrol area. Ideally, there should exist a nonlinear pro ess model
within the whole ight envelope in luding a tuator systems, sensor systems,
engine, time delays, elasti ity, et .:

x_ (t) = f (x(t); u(t)): (11.1)

But the method also works, if there is only a linear state model of the rigid
air raft for one referen e point

x_ (t) = A x(t) + B u(t); (11.2)

as it has been demonstrated in ight tests (se tion 11.2).

119
11.4 Controller Stru ture
Ea h MFC system ontains the main three elements ommand blo k, feedfor-
ward and feedba k ontrollers, gure 11.2.
The ommand blo k ontains the equations to ompute a sele ted state
ve tor x and its time derivative x_ depending on the input signals.

The feedforward ontroller omputes the ontrol inputs whi h are required
for model following. It in ludes an inverse model of the plant. The use of the
state derivative x_ together with x generated by the ommand blo k (whi h
ontains dynami s) allows the use of pure stati gain matri es in the feedforward
ontroller [35.

In pra ti e, no perfe t inversion of the plant to be ontrolled an be pro-

vided, therefore, the feedba k ontroller has to ope with these un ertainties
and disturban e reje tion.

Disturbances
.
xC
Command Feedforward uFF Plant to be
Block Controller + controlled
uFB

xC Feedback x
Controller

Figure 11.2: General layout of a MFC system

11.5 Possible Design Obje tives

The main design obje tive for the omplete MFC system is to full the design
requirements. As already stated, the main three elements ommand blo k,
feedforward and feedba k ontrollers an be designed separately. The design
obje tives for these subtasks an be summarised as follows:

The ommand blo k must ontain the desired dynami behaviour of the
overall MFC system regarding ontrol inputs. Nonlinear ee ts like a tuator
rate and dee tion limitations have to be taken into a ount within the om-
mand blo k.

The feedforward ontroller has to provide an optimum inversion of the plant

to be ontrolled.

The feedba k ontroller has to ensure rapid and smoothly de aying error
dynami s in the presen e of unknown external disturban es and model un er-
tainties in order to maintain the quality of model following. Nonlinear ee ts
(a tuator rate and dee tion limitations) have to be onsidered for its design.

120
11.6 Design Cy le Des ription
The design y le for the MFC approa h is separated into the subtasks for the
ommand blo k, the feedforward, and the feedba k ontrollers.

11.6.1 Command blo k

A pra ti al and simple way to dene the ommand blo k is to use models of
proven systems, su h as an air raft model with Level 1 manual ight ontrol
laws (Fly-by-Wire) or optimum autopilot fun tions. It is obvious that any
ommand hara teristi s implemented in the ommand blo k are limited by
the dynami s of the plant to be ontrolled, mainly be ause of the nonlinear
onstraints of the a tuators. The main onstraint to be onsidered is that the
dynami s of the air raft model implemented in the ommand blo k are not
faster than those of the plant.

11.6.2 Feedforward ontroller

For the design of the feedforward ontroller an inverse model of the plant is
required. Assuming that the plant model an be represented by a linear state
spa e system (equation (11.2)) the following feedforward ontrol law an be
applied [35:

uF F (t) = B 1 [x_ C (t) A xC (t) : (11.3)

This equation indi ates that the inversion does not in lude dynami elements
(whi h means zero order) if the state derivative x_ together with x generated by
the ommand blo k are available. For this pro edure the ontrol input matrix
of the plant B has to be inverted. This leads to the fundamental problem,
that dierential equations des ribing typi al dynami systems to be ontrolled
(air raft, heli opters, industrial robots, et .) often annot be inverted. In most
of these ases the number of ontrol inputs is smaller than the number of states,
therefore, B is a non-square matrix.
One approa h to handle this problem is the appli ation of the Pseudo-
Inverse

 
By = BT Q B BT Q ; (11.4)

whi h strongly depends on the weighting matrix Q. Therefore, an alternative

method is used at DLR Institute of Flight Me hani s:
Assuming that the state ve tor x is of the order n and the input ve tor u
is of the order m (with n > m) equation (11.2) an be written as:
      
x_ 1 = A11 A12 x1 + B1 u (11.5)
x_ 2 A21 A22 x2 B2
The state ve tor to be ontrolled x1 has the order m. Its elements should
be sele ted properly providing that the subsystem of x1 is fully ontrollable.
In this ase B1 is a square matrix and invertible.

121
 In order to de ouple x1 from x2 , the feedforward ontrol law equation (11.3)
is extended to a de oupling term, with x2  x2C :
uF F = B1 1 (x_ 1C A11 x1C ) B1 1 A12 x2C : (11.6)
| {z } | {z }
Inversion De oupling
Dening the ontrol matrixes

M1 = B1 1 ; M2 = M1 A11 ; M3 = M1 A12 ; (11.7)

equation (11.6) an be simplied to:

uF F = M1 x_ 1C + M2 x1C + M3 x2C : (11.8)

This pure linear approa h an be extended to nonlinear elements, if required

[22.

11.6.3 Feedba k Controller

The feedba k ontroller has to ompensate for model un ertainties and dis-
turban e reje tion, while the feedforward ontroller performs the majority of
the ontrol a tivity. The main requirement for the feedba k ontroller an be
dened by:

e(t) = xC (t) x(t) ! Min. : (11.9)

Dierent methods an be utilised to design the feedba k ontroller, su h as

ve tor performan e optimisation [113 or robust ontrol system design methods.
It has been shown that for air raft appli ations a feedba k ontroller using all
signi ant states with proportional and integral terms is su ient. Its stru ture
is illustrated in gure 11.3.

uFB
+

1/s

KP KI
Command Plant
Block states e states
xC - x

Figure 11.3: Stru ture of the feedba k ontroller

The feedba k ontroller is dened by:

Z
uF B (t) = KP e(t) + KI e(t)dt: (11.10)

122
The gains of the feedba k ontroller an be optimised independently from the
layout of the feedforward ontroller and the ommand blo k. A proven pro-
edure used at DLR is based on a numeri al optimisation algorithm [126. A
ve tor ost fun tion allows the formulation of ea h design obje tive separately
and its ombination with individual weighting fa tors k for the ontroller per-
forman e and l for the ontroller a tivity. The formal stru ture of the ost
fun tion is given by:

n
X Z te X m Z t
e2k (t) dt + l u2l (t) dt + :::
e
J = k (11.11)
k=1 0 l=1 0
This ost fun tion has to be tailored to the a tual design problem. For air raft
appli ations mostly a number of about ten gains to be optimised is su ient.
However, for highly elasti air raft this number may in rease.

11.7 A Simple Design Example

In order to demonstrate the design pro ess a very simple example is dened,
gure 11.4: The plant represents a simplied air raft model ontaining only
the short period mode, whi h is represented by the following linear model with
a very poor damping ratio (   0:2):
      
q_ = 0:24 0:016 q + 2:4 :
6:5 t
(11.12)
w_ 80:6 0:67 w
.
qc
u Command M1 q
block qc + t
M2 Plant
+ w

M3

Figure 11.4: Blo k diagram of the design example

The ommand blo k in luding the model to be followed is dened as a rst

order system, whi h provides a ommand for the pit h rate q :
q_ = q u: (11.13)

For the MFC design the matri es M1 to M3 have to be determined a ording

to equation (11.7). In this ase x1 = q and x2 = w is sele ted.
In order to verify this design, the Bode plot from the ommand blo k output
q_C to the plant output q is al ulated, gure 11.5a. The pure integral behaviour
demonstrates that the inversion works orre tly.
The step responses larify the poor damping of the unaugmented plant and
the realised rst order behaviour of the MFC system, gure 11.5b.

123
 In this linear example, any desired dynami behaviour an be implemented
in the ommand blo k, su h as an air raft model with Level 1 Flying Quali-
ties. Under real onditions the nonlinear ee ts of the plant, su h as rate and
dee tion limits should additionally be onsidered in the ommand blo k.

.
a) Bode Plot of frequency response q/qc
Amplitude (dB)
20

-20
Phase (deg)

-90

0.1 1 Frequency (rad/sec) 10

b) Step Responses
0
MFC system response q/u
-1

-2
pure aircraft system response q/ t
-3
0 5 10 15 Time (sec) 20

Figure 11.5: Results in the frequen y and time domain of the design example

124
12. Predi tive Control

Jan M. Ma iejowski1 and Mihai Huzmezan 1

12.1 Introdu tion

Predi tive Control is very dierent from other ontrol te hniques, in that in
its most powerful form it requires the on-line solution of a onstrained opti-
mization problem. This makes it an unlikely andidate for ight ontrol. On
the other hand, it oers some unique benets, sin e it expli itly allows for
hard onstraints, and it an anti ipate pilot ommands if the ight traje tory
is known in advan e. This makes it interesting for ight ontrol, parti ularly
if higher-level ontrol fun tionality is onsidered. In this tutorial hapter we
present the models used by predi tive ontrol, the ontrol problem formula-
tion, dis uss solutions te hniques and ontroller properties, and omment on
the problem of tuning the predi tive ontrol problem formulation so as to meet
design spe i ations.

12.2 General Chara teristi s

Predi tive Control, also known by several other names, su h as Model-Based
Predi tive Control (MBPC), Re eding Horizon Control (RHC), Generalised
Predi tive Control (GPC), Dynami Matrix Control (DMC), Sequential Open-
Loop Optimizing ontrol (SOLO) et , is distinguished from other ontrol method-
ologies by the following three key ideas:

 An expli it `internal model' is used to obtain predi tions of system be-

haviour over some future time interval, assuming some traje tory of on-
trol variables.

 The ontrol variable traje tory is hosen by optimizing some aspe t of

system behaviour over this interval.

 Only an initial segment of the optimized ontrol traje tory is imple-

mented; the whole y le of predi tion and optimization is repeated, typ-
i ally over an interval of the same length. The ne essary omputations
are performed on-line.

The optimization problem solved an in lude onstraints, whi h an be

used to represent equipment limits su h as slew rates and limited authority
ontrol surfa es, and operating/safety limits su h as limits on roll angle, des ent
1
Cambridge University Engineering Dept, Cambridge CB2 1PZ, England

125
rate, et . Predi tive ontrol has hitherto been applied mostly in the pro ess
industries, where the expli it spe i ation of onstraints allows operation loser
to onstraints than standard ontrollers would permit, and hen e operation at
more protable onditions.

The drawba k of this approa h for ight ontrol is of ourse the on-line om-
putational requirement. But this is a temporary problem, whi h will disappear
within a few years as omputing te hnology advan es. If the internal model
is linear, the onstraints are linear inequalities, and the performan e riterion
being optimized is quadrati , then the optimization problem to be solved on-
line is a onvex quadrati program, whi h is a relatively good situation. (See
below for more details.)

Most a ademi publi ations on predi tive ontrol deal with un onstrained
problems. See [174, 27, 222 for some good examples. The usual formulations
then be ome losely related to, or even variants of, the LQ te hnique treated in
hapter 5. In this ase ontrollers an often be pre omputed o-line, but mu h
of the advantage of the predi tive ontrol formulation is lost. In this hapter
we assume that onstraints are an essential part of the problem. The problem
with onstraints is treated in some detail in [197.

12.3 System Models

All ontrol methodologies assume some model of the system being ontrolled.
An unusual feature of predi tive ontrol is that an expli it internal model is
required as part of the ontroller; this internal model may be of the same kind
as the assumed behaviour of the real system, but need not be. For the purposes
of analysing behaviour of the omplete ontrolled system it is usual to assume
the same kind of model, though not ne essarily with the same parameters. (So
the situation is similar to that in Robust Control theory, for instan e  see
hapters 7,8.)

The internal model is required in order to generate predi tions of future

system behaviour, if a parti ular set of future ontrol a tions is assumed. For
this purpose a nonlinear model ould be used, and there have been some studies
of using neural network and other nonlinear models with predi tive ontrol.
Su h models lead to non- onvex optimization problems, however, and to ontrol
s hemes for whi h no analysis is possible. They have not yet been proven to be
useful or ne essary in pra ti e [198. We will therefore assume that the internal
model is linear.

Mu h of the predi tive ontrol literature assumes that a linear time-invariant

model is available in the form of a (multivariable) step or impulse response, and
that predi tions are generated by onvolution: suppose that the multivariable
step response sequen e is given by fgi : i = 0; 1; : : :g, that the ( ontrol) input
ve tor at time k u(k) and that the (to be ontrolled) output ve tor at time
is
k is y(k). Also let
u(k) = u(k) u(k 1) be the hange in the input at time

126
k. Then the output is given by

k
X
y(k) = gk i
u(i) + d(k) (12.1)
i= 1
where it has to be assumed that the open-loop system is asymptoti ally stable
for this to be valid, and d(k) is assumed to be a disturban e a ting on the
output.
In this ase predi tions of the output are omputed by

k+j
X
y^(k + j ) = gk+j i
u(i) + d^(k + j ) (12.2)
i=k+j N

where N is a relatively large integer, and d^(k) is some estimate of d(k + j ).

Frequently the disturban e is estimated as

d^(k) = y(k) y^(k) (12.3)

and it is assumed that future disturban es are the same as the urrent one:

d^(k + j ) = d^(k): (12.4)

The onvolution model is an ine ient one, sin e the same model an be
represented mu h more ompa tly in either transfer fun tion or state-spa e
form. Furthermore, representing the system by a model of this kind removes
the restri tion to stable models. The Generalised Predi tive Control (GPC)
form of predi tive ontrol uses a multivariable transfer fun tion form of model:

A(z 1 )y(k) = B (z 1 )u(k) + n(k): (12.5)

in whi h z 1 is the one-step time delay operator (or the inverse of the z-
transform variable), and A(z 1 ), B (z 1 ) are matri es of polynomials in this
operator  so that [A(z
1 ) 1 B (z 1 ) is the transfer fun tion matrix from the
input ve tor u to the output ve tor y . Although it is not ne essary to asso iate
ea h kind of system model with a spe i disturban e model, it is ommonly
assumed [47 that the disturban e n(k) in this model is generated by passing
white noise through a lter whi h in ludes an integrator:

C (z 1 )
n(k) = e(k) (12.6)
z 1
Inserting an integrator here leads to integral a tion in the ontroller, whi h is
also obtained with the use of onvolution models by the assumption of onstant
future disturban es.
Generating a set of predi tions now involves solving a set of matrix Dio-
phantine equations, but reasonable approximations an be obtained by using
simpler pro edures [47.

127
 The linear model an also be represented in state-spa e form:

x(k + 1) = Ax(k) + Bu(k) + w(k) (12.7)

y(k) = Cx(x) + Du(k) + v(k) (12.8)

where x(k) is the state ve tor and w(k), v(k) are disturban es. For ight on-
trol this model is usually the most appropriate, sin e linearised air raft mod-
els are available in this form, with the state variables representing physi ally
meaningful quantities. If the disturban es are assumed to be sto hasti then
predi tions of the states and outputs an be obtained by using a Kalman lter
[151. If other assumptions are made then some other observer needs to be used
to generate predi tions. To represent sto hasti disturban es with parti ular
spe tra, the state of the model has to be augmented by arti ial states in order
to use a Kalman lter, in just the same way as is done for LQG design [159.
Integral a tion in the ontroller an be obtained by in luding integrators in the
augmented model.

12.4 Problem Formulations

Predi tive ontrol works by hoosing ontrol a tions to minimise some ost
fun tion, su h as

N2
X Nu
X
J (k) = jjM x^(k + ijk) r(k + i)jj2Q + jj
u(k + i)jj2R (12.9)
i=N1 i=1
subje t to onstraints su h as

j
uj (k + i)j  Vj (12.10)

juj (k + i)j  Uj (12.11)

j(M x^)j (k + ijk)j  Xj (12.12)

where x^(k + ijk) is the predi tion of x(k + i) made at time k, M is some
matrix (for example, M =C if only outputs are to appear in J (k)), and r(k)
is some referen e (desired) traje tory for Mx(k). The integers N1 , N2 and
Nu , as well as the weighting matri es Q and R, are in prin iple hosen to
represent some real performan e obje tives (su h as prot maximisation in a
pro ess appli ation [197), but in pra ti e they are tuning parameters for the
ontroller. It is assumed that the ontrol signals are onstant after the end of
the optimisation horizon, namely that
u(k + i) = 0 for i > Nu .
The inequalities an be used to represent a tuator rate limits (12.10), a -
tuator authority limits (12.11), and operating/safety limits (12.12). In these
inequalities uj (k) denotes the j 'th omponent of the ve tor u(k), et , and Vj ,
Uj , Xj are problem-dependent positive values.
The referen e traje tory r(k) an either be the real pilot ommand ve tor
(set-point), or an be generated by passing the pilot ommand through some
lter. In the latter ase the lter design is another tuning parameter. One of

128
the strengths of predi tive ontrol is that if future ommands are known 
for example before the start of a turn or other manoeuvre  then these an
be anti ipated by the ontroller, leading to smoother manoeuvres, fuel savings,
et .

The ost fun tion penalises non-zero hanges
u(k) in the ontrol signals,
rather than the ontrol signals u(k) themselves, sin e the required steady-state
values of u(k ) are not known in advan e. Penalising non-zero u(k ) would `drag'
the ontrol signals away from the required steady-state values, thus preventing
integral a tion, for instan e.

The situation is summarised graphi ally in gure 12.1.

PAST FUTURE SET POINT

REFERENCE PREDICTED OUTPUT

r(k+l)

y(k)=r(k)

u(k+l)
MANIPULATED CONSTANT INPUT
INPUT

k-n k-2 k-1 k k+1 k+l Nu N1 N2

CONTROL HORIZON - Nu

MINIMUM OUTPUT HORIZON - N1

MAXIMUM OUTPUT HORIZON - N2

Figure 12.1: Re eding Horizon Strategy

As was said earlier, ombining a quadrati ost su h as J (k) with linear

inequalities and a linear model leads to a Quadrati Programming (QP) prob-
lem.

Let the solution of this problem be fu (k + i) : i = 1; : : : ; Nu g. Then

u (k + 1) is applied to the system being ontrolled, and the problem is solved
again at time k + 1. (In general one an apply a longer initial segment, and
re-solve the problem at longer intervals.)

Other non-quadrati osts are also possible. For instan e, min-max osts
are sometimes advo ated in order to obtain robust ontrol, while using absolute
values or peak values instead of quadrati fun tions allows the use of Linear
Programming, whi h redu es the on-line solution time [197, 8.

129
12.5 Solution Te hniques

The basi solution te hnique for the onstrained predi tive ontrol problem is
to use a standard QP solver (or LP solver if the ost fun tion is appropriate).
It is important to appre iate that a solution of a QP problem is required on-
line, and that this problem has to be solved in real time. In pro ess ontrol,
where update rates are very low, this is not a big limitation with urrent om-
puting te hnology. (For example, large multivariable problems with a few tens
of inputs, outputs and onstraints, take a few se onds to solve on 486-type
omputers.) But it learly is a problem for ight ontrol, for whi h a speed-up
of something like 103 is required.
There are several possible alternatives to the use of standard QP solvers,
whi h do not seem to have been investigated thoroughly for predi tive ontrol.
The rst is obtained by noting that if there are no onstraints, or if none of
the onstraints is a tive, then the solution an be obtained analyti ally, as the
solution of a linear equation. (For details, see any of the referen es mentioned
previously.) The problem is that one does not know, before omputing the
solution, whether any onstraints are going to be a tive or not. Se ondly, it
is also true that, if one knew exa tly whi h onstraints were a tive, then one
ould again obtain the solution analyti ally. So if one knows that the set of
a tive onstraints at time k is the same as that at time k 1, then one an nd
the solution very qui kly. Furthermore, it will often be the ase that the set of
a tive onstraints an only hange in very limited ways from one step to the
next; it is then feasible to obtain a small set of analyti solutions qui kly, and
he k whi h one is the a tual solution. These approa hes exploit knowledge
and understanding of the parti ular optimisation problem being solved  ie
ight ontrol of a parti ular air raft  whi h a general QP solver annot do.

Another possibility, again not open to a general solver, is to guess that

the solution at time k will be very similar to that at time k 1, and hen e
use that as an initial trial solution. This strategy has been employed in [119,
for example. Of ourse su h a strategy will o asionally go wrong, when a
onstraint is approa hed losely, and a ba k-up pro edure is required for su h
o asions.

One problem with standard QP solvers is that they give up if the optimisa-
tion problem being solved is infeasible, a situation whi h should not o ur with
proper spe i ation, but nevertheless might. (Typi ally infeasibilities o ur
`inside' the algorithm only, and are due to apparently unavoidable onstraint
violations some time in the future; the feedba k a tion of the ontroller usually
restores feasibility before the problem is en ountered by the system.) In [137
the use of Lawson's weighted least-squares algorithm is advo ated, in whi h the
weight is iteratively adjusted to emphasise the most-violated onstraint. This
algorithm solves the QP problem if it an, and gives a `reasonable' solution if
the problem is infeasible.

130
12.6 Controller Properties
When a linear model and quadrati ost is used, the resulting ontroller is
linear time-invariant providing that either no onstraints are a tive, or that a
xed set of onstraints is a tive. (For ea h su h set, a dierent linear ontrol
law results.) Thus the ontrol law an be a linear law for long periods of time.
However, when hard onstraints are approa hed the ontroller an behave in a
very nonlinear way. In parti ular, it may rea t mildly to a disturban e whi h
drives the system away from onstraints, but very sharply to a disturban e
of similar magnitude but in the opposite dire tion, whi h drives the system
towards onstraints.

The ontroller stru ture is very dierent from more onventional ontrollers.
It onsists of a predi tor, whi h an be ompared with onventional ontrollers,
for example by omparing omplexity as measured by the number of state vari-
ables, and an optimiser, whi h annot be ompared in that way. Figure 12.2
shows the stru ture of a predi tive ontroller. Clearly a predi tive ontroller
is more omplex, in terms of behaviour, in terms of algorithm stru ture, and
in terms of omputation y le time, than a onventional ontroller. Veri a-
tion and erti ation is a mu h more formidable task than for a onventional
ontroller.

REFERENCE COMMAND OUTPUT

OPTIMISER PLANT
&
PREDICTOR
using OBSERVER
INTERNAL STATE ESTIMATE

MODEL

Figure 12.2: Stru ture of a Predi tive Controller

It is easy to formulate the predi tive ontrol problem in su h a way that the
ontroller displays (multivariable) integral a tion, and reje ts onstant output
disturban es. This is a hieved by the ombination of a suitable disturban e
model and penalisation of non-zero
u(k) in the ost fun tion rather than of
non-zero u(k). It is not lear, however, how to obtain double-integral (`type 2')
a tion if it is required. An appropriate disturban e model would be required,
but it would also seem ne essary to penalise
2 u(k) =
u(k)
u(k 1)
instead of
u(k) in the ost fun tion. This means that steadily-in reasing on-
trol a tions ould result, whi h would not be a eptable in most appli ations.
Reje tion of persistent but bounded-amplitude output disturban es, su h as
sinusoids, is easily a hieved by in luding a model of the disturban e (in a or-
dan e with the `Internal Model Prin iple') and penalising
u(k).

131
12.7 Design Spe i ations
The problem of translating ontrol system design spe i ations into spe i
hoi es of predi tion and optimisation horizons ( N1 , N2 , Nu ), weighting ma-
tri es ( Q, R), predi tor, and possibly a referen e-generating lter, is a di ult
one and is still a subje t of urrent resear h. Choosing Q, R, and the pre-
di tor is losely related to the hoi e of weighting and ovarian e matri es in
LQG ontrol; there again the relationship between these parameters and the
design spe i ation is very indire t, but experien e gained over many years has
led to some rules of thumb, at least. The problem is made onsiderably more
ompli ated by also having to hoose horizons [222, 173, 119.
If it is assumed that tight ommand-following is attained by the ontroller,
then the hoi e of referen e-generating lter approximately denes the be-
haviour in response to ommands  a kind of model-referen e approa h 
at least for the ase of ina tive onstraints. However, the assumption of tight
model-following may not be realisti .
Time-domain ommand-following spe i ations are, in prin iple, easily a hieved
by formulating appropriate inequality onstraints. For example, restri tions on
overshoot or rise-time during step responses may be formulated as inequality
onstraints. In pra ti e, however, there are problems if too many onstraints
are added, sin e the solution time in reases. One should, however, be wary
of taking responses to parti ular ommands su h as steps to be representative
of behaviour in response to other ommands, sin e the predi tive ontroller is
nonlinear (if onstraints be ome a tive).
Frequen y domain spe i ations an be he ked under the assumption that
no onstraints are a tive, or that a parti ular set of onstraints is a tive. Fre-
quen y response hara teristi s of the ontroller an be omputed (and some
software is available to do this [118) under either assumption, sin e the on-
troller is then linear and time-invariant (assuming a quadrati ost fun tion).
No omplete systemati method is urrently known of modifying the optimi-
sation problem parameters in su h a way as to a hieve parti ular frequen y-
domain hara teristi s, but signi ant progress towards this is reported in [151.
This is parti ularly relevant for a hieving stability and performan e robustness.
Stability of the losed loop is always part of the design spe i ation, even
if only impli itly. In the absen e of a tive onstraints, it is known how to
enfor e stability. Essentially, either the predi tion horizon N2 must be made
large enough, or `terminal' equality onstraints, whi h bind at time k + N2 ,
must be added to the problem formulation. It has been shown that, from
the point of view of stability enfor ement, terminal equality onstraints an
be ex hanged for an innite predi tion horizon [199. Furthermore, several
methods are known of ensuring stability even in the presen e of onstraints,
but under the assumption that the problem posed always remains feasible. This
is a very strong and almost unveriable assumption, and some urrent resear h
is aimed at removing it.
Most stability proofs are based on proving the monotoni ity of the ost
fun tion J (k) with k, and hen e using the ost fun tion as a Lyapunov fun tion.

132
There have also been some attempts at exploiting the pie ewise-linear nature of
the ontroller to prove stability. Whereas obtaining stability is not di ult in
pra ti e for predi tive ontrol s hemes, there are not yet standard pro edures
for obtaining spe ied stability margins. (This is essentially the same problem
as the problem of obtaining spe ied frequen y response hara teristi s, whi h
was dis ussed above.)
Although tuning of predi tive ontrollers remains di ult, mu h progress
is being made, and systemati pro edures, whi h tune only some of the free
parameters, are be oming in reasingly lear [173, 151.

12.8 Appli ations

Predi tive ontrol has mostly been applied in the pro ess industries, and parti -
ularly in the petro hemi al industries. In these appli ations there is strong mo-
tivation to exploit the apability of respe ting onstraints, sin e mu h money is
to be made by operating as lose as possible to onstraints. Also time onstants
are very big, so there is plenty of time to perform the ne essary omputations.
It is important to stress that in these industries predi tive ontrol is a mature
te hnology, whi h is used routinely.
A few papers report the use of predi tive ontrol with high-bandwidth
ele tro-me hani al systems su h as servos and automotive systems [23. Typi-
ally these either restri t the stru ture of the predi tive ontrol law a priori in
order to obtain an easier optimisation problem [5, or pose a problem without
onstraints [68.
Several studies of using predi tive ontrol in aerospa e appli ations have
been reported, though only a minority of these have really addressed the on-
line omputation problem [105, 223, 24, 214, 252.

12.9 Con lusions

Constrained Predi tive Control is radi ally dierent from other ontrol ap-
proa hes whi h are onventionally used, or might be used, for ight ontrol.
Not only is the design method rather dierent, but the implemented algo-
rithm is quite dierent, be ause it works by repeatedly solving an optimisation
problem on-line. As a straight repla ement for those approa hes, it is not ur-
rently ompetitive, primarily be ause of the omputational load. Even when
further advan es in omputing hardware bring the solution time down to a -
eptable levels, whi h they will surely do, the predi tive ontrol approa h will
give greater problems of erti ation than onventional ontrollers, be ause of
the di ulty of analysing the ontroller behaviour.
On the other hand, predi tive ontrol oers some unique benets: its very
nonlinear behaviour when onstraints are approa hed, and its ability to anti -
ipate pilot ommands, instead of merely rea ting to errors propagating round
the feedba k loop.

133
 We believe that predi tive ontrol is worth investigating further for use in
ight ontrol, if its unique benets are exploited to obtain higher-level fun tion-
ality, in addition to routine stability augmentation. This is dis ussed further
in Chapter 25.

A knowledgement
We would like to thank Dr Angel Perez de Madrid, of UNED, for useful om-
ments during the preparation of this hapter.

134
13. Fuzzy Logi Control

Gerard S hram , Uzay Kaymak1

1

and Henk B. Verbruggen1

13.1 Introdu tion

Designing ontrollers for everyday tasks su h as grasping a fragile obje t, driv-
ing a ar, or more ompli ated tasks su h as ying an air raft, ontinue to be
a real hallenge, while human beings perform them relatively easily. Unlike
most onventional ontrol systems, however, humans do not use mathemat-
i al models nor do they use exa t traje tories for ontrolling su h pro esses.
Moreover, many pro esses ontrolled by human operators in industry annot be
automated using onventional, linear ontrol te hniques, sin e the performan e
of these ontrollers is often inferior to that of the operators. One of the reasons
is that linear ontrollers, whi h are ommonly used in onventional ontrol,
may not be appropriate for nonlinear plants. Another reason is that humans
use various kinds of information and a ombination of ontrol strategies, that
annot be easily integrated into an analyti ontrol law. However, a lot of
experien e and knowledge is available from the experts (e.g. the pilot), whi h
an be made expli it and programmed as a ontrol strategy in a omputer.
Knowledge-based (expert) ontrol tries to formalize the domain-spe i
knowledge, and uses reasoning me hanisms for determining the ontrol a -
tion from the knowledge stored in the system and the available measurements.
Knowledge-based ontrol systems try to enhan e the performan e, reliability
and robustness of urrent ontrol systems by in orporating knowledge that
annot be a ommodated in analyti models upon whi h onventional ontrol
algorithms are based. Knowledge-based systems an be used to realize the
losed-loop ontrol a tions dire tly, i.e. repla e onventional losed-loop on-
trollers, or they an omplement and extend onventional ontrol algorithms
via supervision, tuning or s heduling of lo al ontrollers.
A ommon type of knowledge-based ontrol is rule-based ontrol, where
the ontrol a tions orresponding to parti ular onditions of the system are
des ribed in terms of ifthen rules. Fuzzy Logi Controllers (FLCs) are rule-
based systems, where fuzzy sets are used for spe ifying qualitative values of the
ontroller inputs and outputs. Mu h of the expert's knowledge ontains lin-
guisti terms su h as small, large, et ., whi h an be appropriately represented
by fuzzy sets. Using fuzzy logi , experts' (linguisti ) knowledge of the pro ess
ontrol an be implemented. The rst appli ation of fuzzy logi ontrol was in

1
Department of Ele tri al Engineering, Delft University of Te hnology. P.O.Box 5031,
2600 GA Delft, The Netherlands. {g.s hram}{u.kaymak}{h.b.verbruggen}et.tudelft.nl

135
 ement kiln ontrol [114. The rules representing the ontroller a tions were de-
rived from the ement kiln operator's handbook. Sin e then, fuzzy logi ontrol
has been applied to various systems in the hemi al pro ess industry, onsumer
ele troni s, automati train operation, and many other elds [66, 136, 236.
In se tion 13.2, the basi prin iples of fuzzy sets and fuzzy logi are intro-
du ed. Next, the ontrollers are onsidered in detail, followed by a dis ussion
on ontroller tuning in se tion 13.4. In se tion 13.5, software and hardware
tools are des ribed. In se tion 13.6, the possibilities of fuzzy logi for ight
ontrol are dis ussed. The hapter ends with on lusions.

13.2 Basi Prin iples

The basi idea of a fuzzy logi ontroller is to formalize the ontrol proto ol of
a human operator, whi h an be represented as a olle tion of ifthen rules, in
a way tra table for omputers and for mathemati al analysis. As an example,
onsider the ontrol of the F/A-18 during arrier landing [229. Following a
three dimensional ight path, the task involves the ontrol of speed, sink rate,
and angular attitudes in order to allow a safe ship-board landing. The ontrol
strategy of the pilot onsists of several subtasks, e.g. roll angle ontrol. If the
desired roll angle is positive large (roll angle error positive large), then the
pilot imposes a positive lateral displa ement on the sti k:

If roll angle error is positive large

Then lateral sti k position is positive large

The rule des ribes a proportional relation between roll angle error and lat-
eral sti k position. Usually, the rules are a ombination of proportional as well
as derivative a tion in order to redu e rates. A typi al rule from the sink rate
rule base is:

If sink rate error is near zero AND sink a eleration is positive large
Then longitudinal sti k position is negative medium
The rst part of the rules, alled the ante edent, spe ies the onditions
under whi h the rule holds, while the se ond part, alled the onsequent, pre-
s ribes the orresponding ontrol a tion. Both the ante edent and the on-
sequent ontain linguisti terms (large, small, near zero et .) that ree t the
pilot's knowledge of the pro ess. The ante edent ondition is dened as a om-
bination of several individual onditions, using a onne tive, su h as the logi al
AND operation. It is possible that other rules may ombine the ante edent
onditions using dierent onne tives su h as the logi al OR or the omple-
ment NOT. When the rules of the above mentioned type are to be represented
in a form tra table for omputers, one needs to dene the linguisti terms and
the onne tives that operate on the linguisti terms.
In fuzzy ontrol, the linguisti terms are represented by fuzzy sets. Suppose
that the pilot has a general idea of what a small or large value is, without a

136
sharp distin tion. Su h a term an be des ribed by a fuzzy set, represented
by a so- alled membership fun tion [264, whi h is dened on the universe of
dis ourse X as a fun tion:

 : X ! [0; 1:

The position and shape (e.g. triangular, trapezoidal or bell-shaped) of

the membership fun tion depend on the parti ular appli ation. However, in
many ases triangular shapes are preferred be ause they are related to linear
a tions. Consider for example the roll angle error. In Figure 13.1, a number of
triangular-shaped membership fun tions are shown. Note that in this example
the membership fun tions are pairwise overlapping and that their sum is always
1. A roll angle error of 15 degrees belongs for 50% to the set of a positive small
error (PS) and for 50% to the set of a positive medium error (PM). In other
words, the membership degrees P S (15) and P M (15) are both 0.5.

NL NM NS ZE PS PM PL
1

-30 0 15 30

Figure 13.1: Membership fun tions for roll angle error; negative (N), positive
(P), large (L), medium (M), small (S), zero (ZE).

Fuzzy set operations are performed by logi al onne tives su h as AND

( onjun tion), and OR (disjun tion). For this purpose, the logi al onne tives
from onvential Boolean logi have been extended to their fuzzy equivalents.
The generalization of onjun tion to fuzzy sets is done by fun tions alled t-
norms. Disjun tions are generalized by t- onorms. The most ommonly used
onjun tion operators are the minimum and produ t operators. Usually, the
maximum or the probabilisti sum operator is used for the disjun tion. In
Figure 13.2, the onjun tion and disjun tion operations on two fuzzy sets are
shown when the minimum and maximum operators are used, respe tively.

conjunction (AND) disjunction (OR)

Figure 13.2: Conjun tion and disjun tion of two fuzzy sets by minimum and
maximum operator, respe tively.

137
13.3 Fuzzy Logi Control
Using fuzzy sets and fuzzy set operations, it is possible to design a fuzzy reason-
ing system whi h an a t as a ontroller [162. In Figure 13.3, the stru ture of
a typi al fuzzy logi ontroller (FLC) is shown. The ontrol strategy is stored

knowledge base
scaling membership rule membership scaling
factors functions base functions factors

errors control actions

dynamic filter fuzzification reasoning defuzzification dynamic filter
& scaling mechanism & scaling

Figure 13.3: Blo k-s hemati representation of fuzzy logi ontroller.

in the form of ifthen rules in the rule base. They represent an approximate
stati mapping from inputs (e.g. errors) to outputs ( ontrol a tions). The
dynami lters are used to introdu e dynami s, e.g. error and derivative of
error, and to introdu e an integration on the output. Moreover, s aling is per-
formed to keep the signals between the input and output limits for whi h the
fuzzy rules are dened. The membership fun tions provide a smooth interfa e
from the linguisti knowledge to the numeri al pro ess variables. The fuzzi a-
tion module determines the membership degree of the inputs to the ante edent
fuzzy sets. The reasoning me hanism ombines this information with the rule
base and determines the fuzzy output of the rule-based system. In order to
obtain a risp signal, the fuzzy output is defuzzied and s aled.
The omputational me hanism of the FLC an be explained on an example
of a fuzzy variant of a PD (proportionalderivative) ontroller. Simple PD-like
fuzzy ontrol rules an be dened as relations between the ontrol error e, the
error derivative
e and the ontrol a tion u. As an example, assume that the
following two rules are a part of a fuzzy ontroller's rule base:

If e is small and
e is medium Then u is small

If e is medium and
e is big Then u is medium

Triangular membership fun tions are dened for the terms small, medium
and big in the respe tive domains, see Figure 13.4. The omputational me ha-
nism of the FLC pro eeds in ve steps:

1. Fuzzi ation: The membership degrees of the ante edent variables are
omputed ( small (e), medium (e), medium (
e), big (
e)).
2. Degree of fullment: The degree of fullment for the ante edent of ea h
rule is omputed using fuzzy logi operators. The degree of fullment
i determines to whi h degree the ith rule is valid. In the example, the

138
 produ t operator is used:

1 = small (e)
medium (
e)

2 = medium (e)
big (
e):

3. Impli ation: The degree of fullment is used to modify the onsequent

of the orresponding rule a ordingly. This operation represents the if
then impli ation dened as a t-norm, i.e. a onjun tion operator (e.g.
produ t). Hen e, the fuzzy outputs of the rules be ome:

0
1 (u) = 1
small (u)
0
2 (u) = 2
medium (u):

4. Aggregation: the (s aled) onsequents of all rules are ombined into a

single fuzzy set. The aggregation operator depends on the impli ation
fun tion used; for onjun tions, it is a disjun tion operator (e.g. max):

0 0
FLC output (u) = max(1 (u); 2 (u)) 8 u 2 U:
5. Defuzzi ation: the resulting fuzzy set is defuzzied to yield a risp value.
Defuzzi ation an be onsidered as an operator that repla es a fuzzy
set by a representative value. There exists a number of defuzzi ation
methods, su h as the entre of area method. In Figure 13.4, a small arrow
marks the defuzzied value.

The above type of FLC is alled a linguisti ontroller or a Mamdani type

of ontroller. A on eptually dierent type of FLC has been introdu ed by
Takagi and Sugeno [234. In this type of ontroller, the onsequent part is
repla ed by a risp (non)linear fun tion. The ontroller output is omputed
by taking a weighted average of the individual rule outputs. Sin e the rule
outputs are risp, there is no need for defuzzi ation. The ontroller an be
ompared to a gain s heduling ontroller whi h has linear, lo al modules whi h
are smoothly onne ted. In this way, the lo al linear models an be used for
tuning and analysis (e.g. stability) of the FLC, while the global model aptures
the nonlinearity of the system. However, the transparen y of the ontroller is
de reased. In the RCAM design, the Mamdani type is used be ause this type
of ontroller an implement the pilot knowledge most easily (Chapter 26).

Dire t and supervisory ontrol

The motivation for many FLC appli ations is to mimi the ontrol behaviour
of a human operator in a dire t ontrol onguration or in a supervisory ontrol
environment. Many fuzzy logi ontrollers are implemented as dire t ontrollers
in a feedba k loop. As the rule base represents a stati mapping between the
ante edent and the onsequent variables, external dynami lters are used to
introdu e the desired dynami behaviour of the ontroller (Figure 13.3).

139
 2

small medium small
1 1 1 medium( 1
e)
small (e)
1 3
0 0 0
u

big

medium medium
1 1 1
medium(e)
big ( e) 2 3
0 0 0
e e u
product

1 4 max

0
u
5

Figure 13.4: Computational me hanism of a FLC.

A fuzzy inferen e system an also be applied at a higher, supervisory level.

A supervisory ontroller is a se ondary ontroller whi h augments the existing
ontroller for various onditions. Supervisory systems are hara terized by the
additional exibility they bring to the ontrol system. A supervisory ontroller
an adjust the parameters of a low-level ontroller a ording to the pro ess
information, so that dynami behaviour whi h ould not be obtained by the
low-level ontroller due to pro ess nonlinearities or hanges in the operating or
environmental onditions an be a hieved.
An advantage of a supervisory stru ture is that it an be added to already
existing ontrol systems. Hen e, the original ontrollers an always be used as
initial ontrollers for whi h the supervisory ontroller an be used for tuning
the performan e. A supervisory stru ture an be used for implementing dier-
ent ontrol strategies in one single ontroller (heterogeneous ontrol). These
on epts will be shown in Chapter 26, where separate ontrol strategies for low
airspeed and engine failure are in luded.

13.4 Fuzzy Logi Control Design

Two dierent methods an be used for designing fuzzy logi ontrollers:

1. Design the ontroller dire tly from the knowledge available from the do-
main experts.

140
 2. Develop a fuzzy model of the plant from measurements, rst prin iples
and expert knowledge, and use this model to design a ontroller or in or-
porate this model in a model-based ontrol s heme.

The se ond, indire t method is des ribed in e.g. [17, 36, 127. In the rest of
this se tion we will only on entrate on the dire t approa h, whi h will serve
as a guideline for the design in Chapter 26. The design is hara terized by the
following steps:

1. Determine the ontroller inputs and outputs. For this step, one needs
basi knowledge about the hara ter of the plant dynami s (stable, un-
stable, stationary, time-varying, low order, high order, et .), the plant
nonlinearities, the ontrol obje tives and the onstraints. The simplied
plant dynami s together with the basi ontrol obje tives determine the
dynami s of the ontroller, e.g. PI, PD or PID type fuzzy ontroller. In
order to ompensate for the plant nonlinearities, non-stationarity or other
undesired phenomena, variables other than error and its derivative or its
integral may be used as the ontroller inputs. It is, however, important
to realize that with an in reasing number of inputs, the omplexity of the
fuzzy ontroller (i.e. the number of linguisti terms and the total num-
ber of rules) in reases onsiderably. In that ase, rule base simpli ation
and redu tion te hniques need to be used for keeping the number of rules
small [16.

2. Determine the rule base. The onstru tion of the rule base is a ru ial
aspe t of the design, sin e the rule base en odes the ontrol proto ol of
the fuzzy ontroller. Several methods of designing the rule base an be
distinguished. One is based entirely on the expert's intuitive knowledge
and experien e over all operating onditions. Sin e in pra ti e it may be
di ult to extra t all knowledge from the operators, this method is often
ombined with a good understanding of the system's dynami s. Another
method is based on using a fuzzy model of the pro ess from whi h the
fuzzy ontrol rules are derived.

3. Dene the membership fun tions and the s aling fa tors. The designer
must de ide, how many linguisti terms per input variable will be used.
The number of rules needed for dening a omplete rule base in reases
exponentially with the number of linguisti terms per input variable. On
one hand, the number of terms per variable should be low in order to
keep the rule base maintainable. On the other hand, with few terms,
the exibility in the rule base is restri ted with respe t to the a hievable
nonlinearity in the ontrol mapping. The membership fun tions may be
a part of the expert's knowledge, for example the expert knows approx-
imately what a large roll angle error means. If su h knowledge is not
available, membership fun tions of the same shape, uniformly distributed
over the domain, an be used as an initial setting and an be tuned later.
For omputational reasons, triangular and trapezoidal membership fun -
tions are usually preferred to bell-shaped fun tions. Moreover, the latter

141
 fun tions introdu e a nonlinear hara ter whi h may not be desirable in
all ases.

Generally, the input and output variables are dened on losed intervals.
For simpli ation of the ontroller design, implementation and tuning, it
is more onvenient to work with normalized domains, su h as the interval
[ 1; 1. S aling fa tors are used to transform the values from the operat-
ing ranges to these normalized domains. However, one should be aware
that su h s aling fa tors also s ale the nonlinearity in the ontroller whi h
may not always be desirable.

4. Inferen e options. The hoi e of the inferen e operators also inuen es

the shape of the mapping between inputs and outputs. The most used
inferen e method is the max-min method, where the minimum operator
is used for determining the degree of fullment and the impli ation, and
the maximum operator for rule aggregation. Another method is the sum-
produ t inferen e. The latter ombination is useful for an initial, linear
setting of the FLC. This will be explained below.

5. Fine-tuning the ontroller. The implementation of human heuristi s is

formalized by fuzzy logi in a systemati way. Altough ne-tuning the
performan e of the ontroller is essentially a matter of trial-and-error,
an understanding of the inuen e of various parameters an guide the
pro ess. The s aling fa tors, whi h determine the overall gain of the fuzzy
ontroller and also the relative gains of the individual ontroller inputs,
have mainly a global ee t. The ee t of a modi ation of membership
fun tions and rules is more lo alized, for example hanging the onsequent
of an individual rule. The ee t of the hange of the rule onsequent is the
most lo alized and inuen es only that region where the rule's ante edent
holds.

6. Stability analysis. The analysis of the ontroller is mainly based on time

responses. A stability analysis of the nonlinear FLC is in general di ult.
However, results an be obtained by using te hniques from nonlinear
systems theory if a model of the pro ess under ontrol is available [66,
235, 251. The stability is only proven for the parti ular, simplied model.
Re ently, the stability results have also been extended to more general
lasses of systems [43. The resulting ontrollers are usually onservative
be ause of the onservative nature of the stability riteria.

In order to simplify the design, it is possible to initialize the FLC as a linear

fun tion between the input and output bounds. This limits the hoi e of mem-
bership fun tions and operators, and the ontroller be omes easier to analyse.
One way of a hieving linear initialisation is using pairwise overlapping, trian-
gular membership fun tions where the sum of the memberhsip fun tions equals
1. The defuzzied onsequents must be dened su h that the total mapping
of the FLC is a linear fun tion. The defuzzied onsequents are the numer-
i al values after defuzzi ation of ea h individual rule onsequent. Se ondly,

142
produ t operators must be used for determining the degree of fullment and
impli ation. The aggregation and defuzzi ation phase are then ombined in
one step by the so- alled fuzzy-mean method, whereby the FLC output y is
determined as a weighted sum of defuzzied onsequents:

Nr
X
y= i i
i=1
with i and i are the degree of fullment and the defuzzied onsequent of
the ith rule respe tively, and Nr the number of rules. Note that defuzzi a-
tion is performed for ea h individual rule before aggregation takes pla e. In
Chapter 26, the FLCs are initialized in this way.

13.5 Available Hardware and Software Tools

Sin e the development of fuzzy ontrollers relies on intensive intera tion with
the designer, spe ial software tools have been introdu ed by various software
(SW) and hardware (HW) suppliers su h as Omron, Aptronix, Inform, Siemens,
National Semi ondu tors, et . Most of the programs run on a PC under MS-
Windows, although some of them are also available for UNIX systems.
The general stru ture of most software tools is depi ted in Figure 13.5. The

project editor

Figure 13.5: Generi stru ture of a software tool for fuzzy ontroller design.

heart of the user interfa e is a graphi al proje t editor that allows the user
to build a fuzzy ontrol system from basi blo ks. Input and output variables
an be dened and onne ted to the fuzzy reasoning unit. If ne essary, one
an also use pre-pro essing or post-pro essing elements su h as dynami lters,
integrators, dierentiators, et . The fun tions of these blo ks are dened by the
user. The rule base and the related fuzzy sets are dened using the rule base

143
and membership fun tion editors. The rule base editor is a spreadsheet in whi h
the rules an be entered or modied. The membership fun tions editor is used
for dening the shape and position of the membership fun tions graphi ally.
After the rules and membership fun tions are designed, the fun tion of the
fuzzy ontroller an be tested using system analysis and simulation software
(e.g. MATLAB/SIMULINK).
On e the fuzzy ontroller is tested using various analysis tools, it an be
used for ontrolling the plant either dire tly by the environment (via omputer
ports or analog inputs/outputs), or through generating a run-time ode. Most
of the programs generate a standard C- ode and also a ma hine ode for spe i
hardware, su h as mi ro ontrollers or programmable logi ontrollers (PLCs).
An alternative implementation is a multi-dimensional look-up table with a
simple interpolation routine. This ould simplify validation and erti ation in
ase of ight riti al ontrol systems.

13.6 Fuzzy Logi for Flight Control

Re ently, mu h attention has been paid to the appli ation of knowledge-based
ontrol te hniques for ight ontrol [228, 230. It shows that te hniques like
neural networks and fuzzy systems an provide appropriate tools for nonlinear
identi ation [156, 204, ontrol of high performan e air raft [183, 229 (inner
loop as well as outer loop ontrol), heli opters [195, 233, spa e raft [26, 106,
ight ontrol re onguration [142, 182, 263, and advisory systems [111, 232.
In these appli ations, neural networks generally serve as nonlinear, sometimes
adaptive, fun tion approximators, while fuzzy systems are used as supervisory,
expert systems.
An example of a fuzzy logi ontrol appli ation for ight ontrol is [229.
The ne essary knowledge is extra ted from experien ed pilots. In Chapter 26,
pilot heuristi s of ying an air raft are implemented in the design of a FLC as
well. The FLC design onsists of longitudinal and lateral outer loop tra king
ontrollers ombined with lassi al inner loop attitude ontrollers. Additional,
supervisory rules for low airspeed and engine failure are in luded whi h show
how (gain) s heduling and ex eption handling an be readily in orporated.

13.7 Con lusions

A fuzzy logi ontroller an be onsidered from the AI point of view as a
real-time expert system implementing a part of a human operator's or pro ess
engineer's expertise. From the ontrol engineering perspe tive, a FLC is a non-
linear ontroller. Re ently, a lot of resear h eort has been put into fuzzy logi
ontrol. The appli ations in industry are also in reasing. Major produ ers of
onsumer goods use fuzzy logi ontrollers in their designs for onsumer ele -
troni s, dishwashers, washing ma hines, automati ar transmission systems
et . FLC appli ations are beginning to appear in the pro ess industry as well.

144
 One of the main reasons put forward for using fuzzy logi is that an expli it
mathemati al model des ription is not required for the design of a FLC. Instead
the a tions of a human operator, who already has an internal representation
of the plant, are modelled. This an result in a more e ient ontroller
design, saving time and money. This is only true if expli it operator knowledge
is available in a suitable form. Also, for testing and ne-tuning the FLC, a
reasonable simulation model or the pro ess itself should be available. However,
if little experien e or knowledge about the pro ess is present, and it is not
possible to make eld tests for tuning the ontroller, fuzzy logi ontrol may not
be suitable. One has to onsider espe ially the knowledge a quisition bottlene k
if the experts' knowledge is not available expli itly. An alternative is rst
building a fuzzy model of the nonlinear system from measurement data about
the system, and then applying model-based ontrol te hniques.
Many fuzzy logi ontrollers are implemented as dire t ontrollers in a feed-
ba k loop. In situations where an existing ontroller needs to be extended for
several operating onditions or when a more exible ontrol stru ture is re-
quired, supervisory fuzzy ontrol an provide an answer. It is more di ult
to formulate an analyti ontrol law at this level, while a lot of linguisti in-
formation may be available, whi h an be used for designing the FLC. At this
level, the ontrol problem starts to resemble more and more a de ision making
problem, whi h an be solved by te hniques from fuzzy-de ision making.
The implementation of human heuristi s is formalized by fuzzy logi in a
systemati way. This fa t is also re ognized by the industry, and re ently ef-
forts have in reased to dene a European industry standard for the development
methodology of fuzzy logi systems, based on ISO-9000 general system develop-
ment guidelines [248. However, ne-tuning the performan e of the ontroller
is a matter of trial-and-error like in lassi al ontrol, but using the provided
guidelines and an understanding of the inuen e of ontroller parameters, a
satisfa tory ontroller an be obtained.

145
146
 Part II

RCAM part

147
14. The RCAM Design Challenge
Problem Des ription

Paul Lambre hts , Samir Bennani , Gertjan Looye

1 2 2

and Dieter Moormann - 3 4

Abstra t. The RCAM design hallenge problem is dened in this

hapter using two main se tions. The rst se tion dis usses the
basi ight dynami s model, the available inputs, outputs, param-
eters, et . and the modelling of a tuators, disturban es, et . After
that the ontrol design spe i ations are given and the evaluation
pro edure to be performed by all design teams is presented.

14.1 Introdu tion

This hapter provides the RCAM design problem formulation. It is abstra ted
from the GARTEUR FM(AG08) report: Robust Flight Control Design Chal-
lenge Problem Formulation and Manual: the Resear h Civil Air raft Model
(RCAM) [145. This report formed the basis for the RCAM design hallenge,
the results of whi h are given in the hapters 15 through 26. Therefore its
ontents are given here with little modi ation, so that a lear pi ture of the
information that was available to the design hallenge teams is given. However,
sin e the design and evaluation software that was available to the teams is not
supplied with this book, the des ription of this software has been ex luded. It is
remarked that both the software itself and the des ription appeared helpful in
larifying the problem formulation, but was not intended to provide additional
information.
In se tion 14.2, a des ription of the model is given, in whi h analyti al
expressions for all the parameters of interest, states, inputs and outputs of
the system, are derived. A detailed, oje t-oriented des ription of the model
omponents is also in luded (air raft, sensors, a tuators, engines, wind, et .).

1
Hoogovens Corporate Servi es B.V., HR&D-RSP-SDC 3G.16, P.O.box 10000, 1970 CA
IJmuiden, The Netherlands. (Formerly: NLR, Amsterdam.)
2
Delft University of Te hnology (TUD), Fa ulty of Aerospa e Engineering, Kluyverweg 1,
2629 HS Delft, The Netherlands.
3
German Aerospa e Resear h Establishment (DLR), Institute for Roboti s and System
Dynami s Control Design Engineering, Oberpfaenhofen, D-82230 Wessling, Germany.
4
The following authors ontributed to the original RCAM design denition: Pierre Fabre,
Joseph Irvoas, Philippe Mnard (Aerospatiale), Anders Helmersson (LiTH), Jean-Franois
Magni (CERT), Tony Lambre gts (DUT), Alberto Martnez (INTA), Stefano S ala (CIRA),
Phillip Sheen (AVRO), Jan Terlouw (NLR) Hans van der Vaart (TU Delft).

149
In se tion 14.3 the design problem is formulated, and the riteria and pro edure
adopted for evaluation of the proposed design are des ribed.

14.2 Des ription of the RCAM Model

The purpose of this hapter is to dis uss the RCAM model in a general setting,
su h that used nomen lature and terminology an be introdu ed, and some of
the philosophy behind the stru ture and numeri al al ulations in the software
an be highlighted. The hapter is set up to have some tutorial value, but is
by no means omplete in that sense. It is re ommended to onsult a standard
referen e su h as [74 or [35 for more information on the derivation of equations
of motion, et .

14.2.1 Blo k diagram of the system

A six degrees of freedom nonlinear air raft model in luding nonlinear a tuators
(position and rate limited) and a model of disturban es has been proposed by
Arospatiale. A blo k diagram of this model is given in gure 14.1. Ea h
box in this blo k diagram will be overed in more detail in following text. In
subse tion 14.2.3, an analyti al des ription of the air raft dynami s is given. In
subse tions 14.2.4 and 14.2.5, the sensor and a tuator dynami s are detailed.
In subse tion 14.2.6, the analyti al models of wind disturban es are presented.

14.2.2 Nomen lature: inputs, states, outputs, parame-

ters
As far as appli able, nomen lature is used as dened in the Communi ation
Handbook [237. The following tables summarise this nomen lature, as it is
used both for the formulation of the algorithms and the naming of variables
in the software. Additional information an be found in Appendix A of this
do ument.

The inputs to the model are given in table 14.1. In this table, FE denotes the
earth-xed referen e frame, whi h is dened as follows.

The origin OE is lo ated on the runway longitudinal axis at the

threshold. XE is positive pointing towards the north, and we as-
sume that the runway is also dire ted towards the north (runway
00), hen e XE is positive along the runway in the landing dire tion.
Furthermore, ZE is positive downward, and YE is in the appropriate
dire tion for a right handed axis system (positive east).

FB stands for the body-xed referen e frame, whi h is dened as follows.

The origin OB is at the vehi le entre of gravity. XB is positive

forward, ZB is positive downward and YB is positive to the right
(starboard side).

150
 ACTUATOR MODELS
(including nonlinearities)

trim settings uext

controls sim

Specific outputs
U=[ DA for system analysis
feedback uc DT
path DR
THROTTLE1
THROTTLE2]
ACTUATORS lon
Measurements
for longitudinal
control laws

wext
WIND=[ WXE lat
WYE
WZE Measurements
WIND WXB for lateral
WYB control laws
WIND MODELS WZB ]
AIRCRAFT
(constant wind, turbulence,
windshear)
RCAM MODEL
( 6 degrees of freedom,
non linear, landing configuration)

Figure 14.1: Blo k diagram of the system

The three earth-xed wind inputs, u(6)u(8), are intended to be used for on-
stant wind velo ity omponents eg. headwinds, whereas the body-xed wind
inputs, u(9)u(11), are intended to be used for gusts.

The states used internally by the software are expressed in SI units and are
dened in table 14.2. In this table, CoG denotes Centre of Gravity.

The outputs from the model are given in SI units and are shown in table 14.3.
In this table, FV denotes the vehi le- arried verti al frame, whi h is dened as
follows.

The vehi le- arried verti al frame is parallel to the earth-xed ref-
eren e frame but moves with the vehi le. The origin OV is lo ated
at the vehi le's entre of gravity. XV is positive pointing towards
the north, ZV is positive downward, and YV is positive towards the
east.

Only the model outputs labeled as measured an be assumed to be available

as inputs to the ontroller that is to be designed. The simulation outputs
are only intended to be used for evaluation and should not be used in the nal
ontroller. Note that there is some redundan y in the measured signals, e.g.
 an be determined from uV and vV : depending on the ontrol strategy the
most onvenient signals may be used.

151
 Symbol Alphanumeri Name Unit
A DA u(1) = aileron dee tion rad
T DT u(2) = tailplane dee tion rad
R DR u(3) = rudder dee tion rad
T H1 THROTTLE1 u(4) = throttle position of engine 1 rad
T H2 THROTTLE2 u(5) = throttle position of engine 2 rad
W xE WXE u(6) = Wind velo ity in the x-axis of FE m/s
W yE WYE u(7) = Wind velo ity in the y-axis of FE m/s
W zE WZE u(8) = Wind velo ity in the z-axis of FE m/s
W xB WXB u(9) = Wind velo ity in the x-axis of FB m/s
W yB WYB u(10) = Wind velo ity in the y-axis of FB m/s
W zB WZB u(11) = Wind velo ity in the z-axis of FB m/s

Table 14.1: Input denitions

Symbol Alphanumeri Name Unit

p P x(1) = roll rate (in FB ) rad/s
q Q x(2) = pit h rate (in FB ) rad/s
r R x(3) = yaw rate (in FB ) rad/s
 PHI x(4) = roll angle (Euler angle) rad
 THETA x(5) = pit h angle (Euler angle) rad
PSI x(6) = heading angle (Euler angle) rad
uB UB x(7) = x omponent of inertial velo ity in FB m/s
vB VB x(8) = y omponent of inertial velo ity in FB m/s
wB WB x(9) = z omponent of inertial velo ity in FB m/s
x X x(10) = x position of air raft CoG in FE m
y Y x(11) = y position of air raft CoG in FE m
z Z x(12) = z position of air raft CoG in FE m

Table 14.2: State denitions

Usually, it is possible to dene geometri air raft parameters within the body-
xed referen e frame. However, in the ase of RCAM this is not allowed, as
the CoG is not a geometri ally xed point. For this reason, a measurement
referen e frame FM is dened.

The measurement referen e frame is geometri ally xed to the air-

raft. The origin OM is lo ated at the leading edge of the mean
aerodynami hord, whi h is denoted as . XM is positive pointing
ba kwards, YM is positive pointing to the right (starboard), and
ZM is positive pointing up.

It is assumed that the aerodynami entre of the wing-body onguration

(ACwb ) is also geometri ally xed: its o-ordinates in FM are ( 0:12 ; 0 ; 0).
With these denitions, it is now possible to spe ify the parameters used in
RCAM: they are given in table 14.4.

Finally, RCAM provides the possibility to study the ee t of the parameter
hanges dened in table 14.5.

152
 Symbol Alphanumeri Name Unit
Measured
q Q y(1) = pit h rate (in FB ) = x(2) rad/s
Fx 
nx NX y(2) = horizontal load fa tor (in FB ) = mg -
nz NZ y(3) = verti al load fa tor (in FB ) = mg 1
Fz
-
wV WV y(4) = z omponent of inertial velo ity in FV m/s
z Z y(5) = z position of air raft CoG in FE = x(12) m
VA VA y(6) = air speed m/s
V V y(7) = total inertial velo ity m/s
 BETA y(8) = angle of sideslip rad
p P y(9) = roll rate (in FB ) = x(1) rad/s
r R y(10) = yaw rate (in FB ) = x(3) rad/s
 PHI y(11) = roll angle (Euler angle) = x(4) rad
uV UV y(12) = x omponent of inertial velo ity in FV m/s
vV VV y(13) = y omponent of inertial velo ity in FV m/s
y Y y(14) = y position of air raft CoG in FE = x(11) m
 CHI y(15) = inertial tra k angle rad
Simulation
PSI y(16) = heading angle (Euler angle) = x(6) rad
 THETA y(17) = pit h angle (Euler angle) = x(5) rad
 ALPHA y(18) = angle of atta k rad
GAMMA y(19) = inertial ight path angle rad
x X y(20) = x position of air raft CoG in FE= x(10) m
ny NY y(21) = lateral load fa tor (in FB )= Fy -
 see equations 14.1 and 14.5 mg

Table 14.3: Output denitions

14.2.3 Air raft dynami s model

This subse tion des ribes the RCAM dynami s model orresponding to the
AIRCRAFT blo k in gure 14.1. The dynami obje ts are depi ted in gure 14.2.

These obje ts are:

 body des ribes the body dierential equations of motion (see subse -
tion 14.2.3);

 two transformation obje ts des ribe the o-ordinate transformation be-

tween the body-xed o-ordinates of the body obje t and the geodeti o-
ordinates of the gravity obje t, and between the body-xed o-ordinates
of body and the geodeti o-ordinates of wind, respe tively (see subse -
tion 14.2.3);

 al airspeed des ribes the relationship between the inertial movement, the
wind, and the movement relative to the air (see subse tion 14.2.3);

 engine_1 and engine_2 des ribe the relevant engine behaviour (see sub-
se tion 14.2.3);

 atmosphere des ribes the atmosphere model (see subse tion 14.2.3);

 aerodynami des ribes the aerodynami for es and moments (see subse -
tion 14.2.3);

153
 Symbol Alphanumeri Name Default Unit
m MASS = air raft total mass 120 000 kg
Aerodynami Parameters
 CBAR = mean aerodynami hord 6.6 m
lt LTAIL = distan e between AC of the wing-body 24.8 m
(ACwb ), and AC of the tail (ACt )
S S = wing planform area 260.0 m2
St STAIL = tail planform area 64.0 m2

x DELX = x position of the CoG in FM 0.23  m

y DELY = y position of the CoG in FM 0 m

z DELZ = z position of the CoG in FM 0 m
Engine Parameters
XAP T 1 XAPT1 = x position of appli ation point of 0.0 m
thrust of engine 1 in FM
YAP T 1 YAPT1 = y position of appli ation point of 7:94 m
thrust of engine 1 in FM
ZAP T 1 ZAPT1 = z position of appli ation point of 1:9 m
thrust of engine 1 in FM
XAP T 2 XAPT2 = x position of appli ation point of 0.0 m
thrust of engine 2 in FM
YAP T 2 YAPT2 = y position of appli ation point of 7:94 m
thrust of engine 2 in FM
ZAP T 2 ZAPT2 = z position of appli ation point of 1:9 m
thrust of engine 2 in FM

Table 14.4: Parameter denitions

Parameter Bounds
m MASS : 100 000 kg < m < 150 000 kg

x DELX : 0.15  <
x < 0.31 

y DELY : 0.03  <
y < 0.03 

z DELZ : 0.0  <
z < 0.21 

Table 14.5: Possible parameter hoi es in RCAM

 gravity des ribes the gravitational inuen e (see subse tion 14.2.3).

Body equations of motion

The following two subse tions give the equations of motion for a rigid body
with 6 degrees of freedom and other air raft motion relevant equations whi h
as used within RCAM. For a more detailed derivation and explanation of these
equations a referen e su h as [74 or [35 should be onsulted.

Translational motion. The equations for the translational movement in body-

xed o-ordinates are derived from the for e equation,

F = m ( aB + !  VB ) (14.1)

F is the sum of for es due to the engines, the aerodynami s and gravity, m
is the mass of the air raft, VB is the inertial velo ity and ! is the rotational
velo ity expressed in body-xed o-ordinates. The a eleration (in body-xed

154
 RCAM aero
u
sim
body6DOF
RCAM equations

engine1 COG
airspeed

kinetic
RCAM air

wind
long
bodyfixed bodyfixed
engine2
Trafo Trafo
veh.carried veh.carried

const. gravity const.

lat
Earth
atmosphere wind gust

wind
gust

Figure 14.2: Dynami obje ts of RCAM air raft model inside the AIRCRAFT
blo k of gure 14.1. Conne tion arrows between obje ts hara terise physi al
intera tions

system) is the time derivative of velo ity:

2 3
dV d uB
aB = B = 4 vB 5 (14.2)
dt dt w
B
and the velo ity is the time derivative of the position ve tor expressed in the
vehi le- arried verti al frame:
2 3
d XV d x
VV = = 4y5 (14.3)
dt dt z
Additionally, some air raft spe i quantities are dened as follows:
The height h, whi h is the negative z - o-ordinate in the vehi le arried system
h= z (14.4)

The load fa tor n is dened as the relation of the external for es F (equa-
tion 14.1) to the gravity for e mg , with all quantities given in the body-xed
oordinate system. In order to have a zero load fa tor for horizontal ight, the
z- omponent of n is redu ed by one.
2 3 2 3
nx F 0
n = 4 ny 5 = 405 (14.5)
nz mg 1

155
The inertial ight path angle, , is given as a fun tion of the speed omponents
in the vehi le- arried verti al referen e frame

tan = p 2wV 2 (14.6)

uV + vV
The tra k angle, , is also dened as a fun tion of the the speed omponents
in the vehi le- arried verti al referen e frame

v
tan  = V (14.7)
uV
Rotational motion. The equations of motion for the rotational movement of a
rigid body in the body-xed axis system are derived from the moment equation,

M = I !_ + !  I! (14.8)

M is the sum of moments about the entre of gravity due to the engines
and the aerodynami s, ! is the inertial rotational velo ity, and !_ is the iner-
tial rotational a eleration in the body-xed axis system. Using the standard
notation [74 we get:

2 3 2 3
p_ d 4p5
!_ = q_ =
4 5 q (14.9)
r_ dt r
Again using standard notation [74, the relation between the rotational
velo ities and the Euler angles is;

2 3 2 32 3
d 4 _ 1 sin  tan  os  tan  p
= _ 5=4 0 os  sin  5 4 q 5 (14.10)
dt _ 0 sin = os  os = os  r
1
For a normal air raft , the inertia tensor I dened in the body-axis frame
is;
2 3 2 3
Ix 0 Ixz 40:07 0 2:0923
I = 4 0 Iy 0 5 = m4 0 64 0 5 (14.11)
Ixz 0 Iz 2:0923 0 99:92
where all numbers are expressed in square metres, m .
2

Co-ordinate transformation (Body-Fixed , Vehi le-Carried)

The rotations between the body-xed and the vehi le- arried o-ordinate sys-
tem are depi ted in gure 14.3. To des ribe the angular orientation of the
air raft, a transformation using the three Euler angles ,  , and is ne es-
sary. This transformation is a hieved by initially rotating the vehi le- arried
verti al system about the ZV -axis by the heading angle . Then, the result
is rotated about the k2 -axis by the pit h angle , and nally the body-xed
1
Normal air raft are assumed symmetri about the OXZ body axis plane.

156
 xB

xV
k1

yV

xV, yV k2
xB , z v
yB

yB , z B
zB

k3

zV

Figure 14.3: Co-ordinate transformation body-xed , vehi le- arried

referen e frame is obtained by rotating the result of that by the roll angle 
about the XB -axis. This results in the following transformation matrix from
the vehi le- arried verti al axis system to the body-xed axis system:

RBV =
2 32 32 3
1 0 0 os  0 sin  os sin 0 (14.12)
4 0 os  sin  5 4 0 1 0 5 4 sin os 05
0 sin  os  sin  0 os  0 0 1

Note that RBV = RVT B .

For example, the transformation of velo ities from the vehi le- arried verti al
frame FV to the body-xed referen e frame FB is given by:

VB = RBV VV (14.13)

with
2 3 2 3
uB uV
VB = 4 vB 5 and VV = 4 vV 5 (14.14)
wB wV
Similarly, the a elerations, rotational velo ities, positions, for es and moments
an be transformed between the o-ordinate systems.

157
Cal ulation of airspeed
The ve tor airspeed, Va is the dieren e between the inertial velo ity of the
air raft, VB , and the wind velo ities, WB and WE (see table 14.1).
Expressed in the body-xed o-ordinate system this is al ulated as:

Va = VB WB RBV WE (14.15)

Hen e, with
2 3
ua
Va = 4 va 5 (14.16)
wa
the airspeed VA is given as:
p
VA = (ua 2 + va 2 + wa 2 ) (14.17)

Next, the angle of atta k, , and the angle of sideslip,  , are dened as:
w
tan  = a (14.18)
ua
v
sin  = a (14.19)
VA
The derivatives of  and  with respe t to time are:

a u aax wa
_ = az a2
ua + wa 2
(14.20)

a ( u 2 + wa 2 ) va ( aax ua + aaz wa )
_ = ay a p
VA 2 ua 2 + wa 2
(14.21)

where aax , aay , and aaz are the x, y , and z -time derivatives of the airspeed in
body-xed o-ordinates. (e.g. aax =
dua ).
dt

Aerodynami equations
The equations dening aerodynami for es and moments are determined by
means of aerodynami oe ients. Depending on the method of modelling
these oe ients may be dened in dierent referen e frames; e.g. FW , FS , or
FB . The referen e frame for aerodynami for es and moments that is used in
RCAM is the stability axis frame FS .
Aerodynami for es. The aerodynami for es are determined by means of aero-
dynami oe ients for drag, sidefor e and lift ( CD , CY , CL ), whi h are given
as fun tions of the angle of atta k, , the sideslip angle, , and the ontrol
surfa e dee tions.

The aerodynami lift oe ient, CL , is dened as (see gure 14.4);

CL = CLwb + CLt (14.22)

158
C Lwb is the lift oe ient of the wing and body. It a ts on the aerodynami
entre of the wing and body and is parallel with the Z
S axis. It is only a
fun tion of angle of atta k , and for  < 19 degree it is given by the following
equations:

5:5 ( 0 )   14:5 180
 rad
CLwb = 3 2  rad (14.23)
768:5 + 609:2 155:2 + 15:2  > 14:5 180
Here, 0 is the angle of atta k at whi h the wing/body lift is zero:


0 = 11:5 (14.24)
180
The maximum lift oe ient is obtained at an angle of atta k of  = 18 degree.
Negle ting the ee t of the tailplane, this is al ulated from equation 14.23 as:


CLmax = CLwb ( = 18 ) = 2:75 (14.25)
180
The lift oe ient of the tailplane, CLt a ts on the aerodynami entre of
the tailplane and is also parallel with the ZS axis. It is given as:
S
CLt = t 3:1 t (14.26)
S
where t denotes the angle of atta k of the tailplane and is al ulated from the
following equations:

q lt
t =  " + T + 1:3
VA
d
" = ( 0 ) (14.27)
d
d
= 0:25
d
Here " is the downwash angle, T is the tailplane dee tion, q is the air raft
pit h rate, and lt is the longitudinal distan e between the aerodynami en-
tre of the tailplane and the aerodynami entre of the wing and body. (See
gure 14.4).
The aerodynami drag oe ient, CD , is a fun tion of the angle of atta k
; drag of the tailplane is negle ted and it is assumed that CD a ts on the
aerodynami entre of wing and body:

CD = 0:13 + 0:07
(5:5 + 0:654)2 (14.28)

The aerodynami sidefor e oe ient, CY , is also assumed to a t on the aero-

dynami entre of wing and body and is given by the following equation:

CY = 1:6 + 0:24R (14.29)

where  is the angle of sideslip and R is the rudder dee tion.

These non-dimensional oe ients an now be onverted to dimensional for es

using the following relationships:

159
 CL
wb

xB
CoG CL
CD t

Va T
AC
AC t

q=0 lt

Figure 14.4: Illustration of aerodynami for es

 Aerodynami for e along XS

1
X = D = CD V 2S
2 A
(14.30)

 Aerodynami for e along YS

1
Y = CY V 2S
2 A
(14.31)

 Aerodynami for e along ZS

1
Z = L = CL  VA 2 S (14.32)
2

To al ulate the translational motion of the air raft using equation 14.1,
these for es need to be resolved into body axis for e omponents. The resolution
from stability axes for es, (D; Y; L), into the body-axes for es, (FxA , FyA , FzA ),
is given by the following expressions:

FxA = L sin  D os 
FyA = Y (14.33)

FzA = L os  D sin 

Aerodynami moments. The moments due to the air raft aerodynami s are
determined by means of the moment oe ients, ( Cl ; Cm ; Cn ), whi h are as-
sumed to a t about the aerodynami entre of the wing and body and are given

160
by the following equation:
2 3 2 3
Cl 1:4 
4 Cm 5 = 4 0:59 3:1 SSt lt ( ) 5
Cn (1  15 180 )
2  3 2 3
11 0 5 p
0 5 V A 4 q 5
6 7
+ 4:03 SSt l2t
2
4 0 (14.34)

2
1:7 0 11:53 2 3r
0:6 0 0:22 A
+ 4 0 3:1 SSt lt 0 5 4 T 5
0 0 0:63 R
where
p, q,and r are the rotational rates in body axes,
A is the aileron dee tion,
T is the tailplane dee tion,
R is the rudder dee tion.
The moment oe ients about the entre of gravity are al ulated from these
aerodynami entre based oe ients using the following equation:
2 3 2 3 0 1 0 2 31
ClCG Cl
x 0:12 CD
4 CmCG 5 = 4 Cm 5 + 
y A  RBS
4 CY 5A
CnCG 2
C3n 02
z 32
CL 31 (14.35)
Cl 0 CZ CY
x 0:12
= 4 Cm 5 + 4 CZ 0 CX 5 4
y 5A
Cn CY CX 0
z
with
2 3
os  0 sin 
RBS = 4 0 1 0 5 CX = CD os  + CL sin  (14.36)
CZ = CL os  CD sin 
sin  0 os 
The following expressions are used to onvert these non-dimensional mo-
ments oe ients into dimensional moments:

 Rolling moment in body axes

1
LA = ClCG  V 2 S 
2 A
(14.37)

 Pit hing moment in body axes

1
MA = CmCG  VA 2 S  (14.38)
2
 Yawing moment in body axes

1
NA = CnCG  V 2 S 
2 A
(14.39)

161
These moments, in ombination with the moments due to thrust are then used
to al ulate the rotational motion of the air raft from equation 14.8.

RCAM engine thrust al ulation

The RCAM is a twin engined air raft model, and the thrust provided by ea h
of the two engines is assumed to be aligned with the x-body axis.
The thrust produ ed by a single engine is given by

Fi = T Hi mg;
i = 1; 2 (14.40)

with m at the nominal mass of 120,000 kg and T H1 and T H2 determined by

the setting of the throttle handles. In equation 14.40, T Hi should be expressed
in radians: this has no physi al meaning but appears to be onvenient in the
al ulations. The allowed value of T Hi lies between

0:5 180 and

10 180 radians.
Note that the maximum thrust to weight ratio is about 0.35 (for both engines
together). Hen e, the engine thrust ve tor at the enter of gravity is given in
FB as:
2 3
F1 + F2
Fp = 4 0 5 (14.41)
0
Due to the geometri lo ation of the engines, see gure 14.5, the engine
thrusts also ontribute to the moments a ting on the air raft. These moments

zM
6 6
front

F1 F2
6 6
O M - yM  front
- xM
P1 P2 ACwb

ACwb

F1 ; F2
P1 ; P 2

?xM
Figure 14.5: Appli ation points of thrusts.
P1 and P2 are the points where the thrust is applied.

an be al ulated about the entre of gravity as follows:

2 3 2 3

x XAP T i Fi
TEi = 4 YAP T i
y 5  4 0 5 (i = 1; 2) (14.42)

z ZAP T i 0
where
x,
y ,
z , XAP T i , YAP T i and ZAP T i are dened in table 14.4.

162
Atmosphere
The atmosphere is onsidered to be onstant, irrespe tive of height and posi-
tion, and onsequently we an dene the following:

kg
 = 1:225 3
m
N
P = 101325:0 2 (14.43)
m
T = 288:15 K
where  is the density of air, P is the stati air pressure, and T is the absolute
temperature.

Gravity model
Due to the restri ted altitude range to be used with this model, gravity is not
onsidered to be a fun tion of altitude. Hen e, gravity is assumed to have a
onstant value of:

g = 9:81 m/s2 (14.44)

14.2.4 Sensor models

Models are not provided for the hara teristi s of the sensors: they are all
assumed to be perfe t.

14.2.5 A tuator models and engine dynami s

Both a tuators and engines are assumed to have rst order system dynami s
with rate limits and saturations. The time onstants of the rst order system
dynami s are:

 engine models: 1.5 s,

 ailerons and tailplane a tuators: 0.15 s, and

 rudder a tuator: 0.3 s.

Numeri al values for rate limits and saturations are given as follows.

 Rate limits for throttle movement are:

rising slew rate

= 1:6 180 = 1:6 180
rad/s, falling slew rate
 rad/s,

 throttle limits (saturations) are:

 rad 
0:5 180 T Hi  10 180 rad.


In ase of engine failure we an assume that the throttle setting for the
failed engine redu es to
 rad
T Hi = 0:5 180 with rst order system dynami s
given by the transfer fun tion 1=(1 + 3:3s).

163
    _  25  rad/s;
25 180
180 rad,
Rate limits for aileron dee tion are: A
saturations of aileron dee tion are: 25 180  A  25 180


 rate limits for tailplane dee tion are:

  _  15  rad/s;
15 180 T 180
saturations of tailplane dee tion are: 25 180  T  10 180
  rad,

 rate limits for rudder dee tion are:

  _  25  rad/s;
25 180 R 180
saturations of rudder dee tion are: 30 180  R  30 180
  rad.

14.2.6 Wind turbulen e model

Turbulen e is a sto hasti pro ess that an be dened by velo ity spe tra. Com-
monly used velo ity spe tra for turbulen e modelling are the Dryden spe tra.
For an air raft ying at a speed V through a frozen turbulen e eld with a
spatial frequen y of
rad/m, the ir ular frequen y of the turbulen e an be
al ulated as:
! =V

rad/s (14.45)

With this, the spe tra an be des ribed as follows:

2Lug 1
ug (!) = u2g
V (1 + (Lug V! )2 )
L 1 + 3(Lvg V! )2
vg (!) = v2g vg
V (1 + (Lvg V! )2 )2
(14.46)

L 1 + 3(Lwg V! )2
wg (!) = w2 g wg
V (1 + (Lwg V! )2 )2
The turbulen e s ale lengths L L L
ug , vg , wg and turbulen e standard de-
viations ug , vg , wg are dependent on altitude and atmospheri onditions.
As an indi ator for the atmospheri onditions it is possible to take the wind
speed at 20 ft above the ground ( W20 ). For moderat onditions, W20 = 15:4
m/s (30 kts) is sele ted. The turbulen e standard deviation wg is then given
as follows:
wg = 0:1W20 (14.47)

ug and vg are assumed to be fun tions of wg and the altitude h.
For h < 305 m (1000 ft):

wg
ug = vg =
(0:177 + 0:0027h)0:4
(14.48)

and for h > 305 m (1000 ft):

ug = vg = wg (14.49)

The turbulen e s ale lengths Lug , Lvg and Lwg are al ulated as a fun tion
of altitude: for 3 < h < 305 m:
h
Lug = Lvg =
(0:177 + 0:0027h)1:2
(14.50)

164
 Lwg = h (14.51)

and for h > 305 m we take:

Lug = Lvg = Lwg = 305 m (14.52)

With this pro edure, the gust velo ities ug , vg and wg are dened in the stability
referen e frame. However, as an approximation the RCAM inputs W xB , W yB
and W zB are used.

To simulate turbulen e, white noise is ltered through forming lters. These

lters an be derived from the Dryden spe tra given in equation 14.46. As an
example, the transfer fun tion of the lter for simulating the gust velo ity wg
will be onsidered.
Given white noise w, the spe trum of wg an be obtained as:

wg = jHwg w (!)j2 w (14.53)

Where w = 1, and Hwg w (!) is the frequen y response fun tion of the forming
lter. Therefore,

L 1 + 3(Lwg V! )2
w2 g wg = jHwg w (!)j2 = Hwg w (!)Hwg w ( !)
V (1 + (Lwg V! )2 )2
(14.54)

To obtain a stable and minimum phase lter, the following frequen y re-
sponse fun tion is sele ted:

r p
Lwg 1 + 3 LVwg j!
Hwg w (!) = wg (14.55)
V (1 + LVwg j!)2

Repla ing the variable j! by s, the following transfer fun tion is obtained:
r p
Lwg 1 + 3 LVwg s
Hwg w (s) = wg (14.56)
V (1 + LVwg s)2

The transfer fun tion for generating vg is equivalent.

The transfer fun tion for generating ug an be found as:

r
2Lug 1
Hug w (s) = ug (14.57)
V 1 + Lug s
V

It is important to note that for orre t appli ation of these lters the white
noise inputs need to be independent.
For a more detailed dis ussion on turbulen e modelling, the reader is re-
ferred to for example [35.

165
14.3 Design Problem Formulation and Evalua-
tion Criteria
14.3.1 Motivation design and evaluation riteria
Within the aerospa e industry there is a large amount of experien e in the
ight ontrol system design area. For this reason, the main obje tive of the
ontrol problem stated here is not so mu h to obtain a satisfa tory ontroller,
but more spe i ally to exhibit approa hes whi h might redu e the omplexity
of ontrol laws and the overall ontrol system design y le.
Some of the main features addressed by modern ontrol design te hniques
provide the possibility to take into a ount:

 the multivariable nature of the ontrol problem

 the non linear behaviour of the plant

 the time-varying nature of the plant

 robustness to parameter hanges and un ertainties

 simultaneous performan e and robustness spe i ations.

From the onsideration of these features it is expe ted that improvements ould
be made in areas su h as:

 ontrol system ar hite ture development

 ontrol law design y le

 ontrol design solution

 ontrol system implementation

The RCAM design hallenge onsists of the synthesis of a ontrol law apable
of fullling an approa h to landing under various external onditions eg. turbu-
len e and windshear, while being robust to parameter hanges. Furthermore,
the air raft guidan e must not degrade under engine failure. Details on the
design obje tives are given in subse tion 14.3.2.
For the uniform omparison of all design entries from the design hallenge
parti ipants, a set of evaluation riteria is formulated in subse tion 14.3.3. To
evaluate proper ontrol system logi and to make the hallenge more realisti ,
an evaluation traje tory has been designed to ree t typi al phases during
approa h to landing. The evaluation riteria given in this subse tion are based
on sets of signals from whi h ertain hara teristi s will be al ulated. All
designs should be able to tra k the given traje tory within the spe ied bounds.
Note that the hoi e of a traje tory as an evaluation riterion is independent
of the ontrol law and ontrol design methodology.
An important subje t onsidered in this hapter is the translation of design
obje tives into evaluation riteria: the evaluation riteria should be su iently

166
representative for the onsidered design obje tives, but will not be able to
over all aspe ts. It is asked that the ben hmark problem parti ipants onsider
the design obje tives given in subse tion 14.3.2 and for them to use their own
methods to illustrate to what extent these are met by their ontroller design.
For instan e, we give robustness spe i ations in terms of real parameter vari-
ations, although they are often also onsidered in the frequen y domain or in
terms of gain and phase margins. The evaluation pro edure is only aimed at
obtaining an obje tive measure for omparison with other designs.

14.3.2 Design riteria

Introdu tion
The ontroller design problem for the RCAM model is hara terised by a num-
ber of fundamental trade-os between oni ting design spe i ations. For
typi al air raft autopilot systems we re ognise ve lasses of riteria:

 performan e riteria: these ree t tra king error and disturban e reje -
tion hara teristi s of ertain signals;

 robustness riteria: these ree t the stability bounds with respe t to

parameter variations;

 ride quality riteria: these ree t the desire to obtain su ient passenger
and pilot omfort in the form of bounds on ertain maximum allowable
a elerations and minimum damping levels;

 safety riteria: these ree t envelope safeguards;

 ontrol a tivity riteria: these are a measure of the power onsumed by

the ontrols and also give an indi ation of fatigue ee ts.

Performan e riteria
The performan e of the ontrolled system an be spe ied in terms of ommand
response hara teristi s to normalised referen e signals, tra king error and dis-
turban e reje tion features (see [132). The ommand response hara teristi s
are dened in terms of rise time tr , settling time ts and overshoot Mp . Rise
time is dened here as the time the unit step response y(t) takes from y = 0:10
toy = 0:90, i.e., tr = t(y90% ) t(y10%). Settling time is here dened as the time
fory(t) to a hieve 99 per ent of its nal value. Finally, overshoot is dened as
the relative peak of y (t), i.e., Mp =
(ypeak y(1))  100% (see [82).
y(1)
P1- Lateral deviation. The ontrolled air raft's lateral deviation, eyb (t), dened
as the dieren e between the a tual and ommanded lateral air raft position,
y(t) y (t), should be redu ed to 10 per ent within 30 s.
There should be very little overshoot in the response to a unit step in
lateral ommand signals at altitudes above 305 m (1000 ft), i.e., Mp < 5%.
At lower altitudes Mp may in rease to 30% in order to obtain higher tra king

167
performan e. There should be no steady state error due to onstant lateral
wind disturban es. In the nal phase of ight (landing approa h glide path)
the lateral deviation from the desired ight path should not ex eed that given
in gure 14.6.

6Maximum deviation from lo alizer path 6Maximum verti al deviation

20 m 6m

5m 1.5 m
-
Altitude
-
Altitude
100 ft 400 ft 100 ft 400 ft

Figure 14.6: Maximum lateral de- Figure 14.7: Maximum verti al de-
viation viation

P2- Altitude response. The ontrolled system should be able to tra k altitude
ommands, h , with rise time tr < 12 s and settling time ts < 45 s. There
should be very little overshoot in the response to unit steps in altitude om-
mands at altitudes above 305 m (1000 ft), i.e., Mp < 5%. At lower altitudes
Mp may in rease to 30% in order to obtain higher tra king performan e. In the
nal phase of ight (landing approa h glide path) the verti al deviation from
the desired ight path should not ex eed that given in gure 14.7

P3- Heading angle response. The ommanded heading angle, , should be

tra ked by the a tual heading angle, , with a rise time tr < 10 s and and
settling time ts < 30 s. There should be very little overshoot in the response
to unit steps in heading ommands at altitudes above 305 m (1000 ft), i.e.,
Mp < 5%. At lower altitudes Mp may in rease to 30% in order to obtain
higher tra king performan e. For unit RMS intensity lateral Dryden gust, the
RMS of the heading angle error in losed loop should be less than that in open
loop.

P4- Flight path angle response. The ommanded ight path angle, , should
be tra ked by the a tual ight path angle, , with a rise time tr < 5 s and
settling time ts < 20 s. There should be very little overshoot in the response
to unit steps in ight path angle ommands at altitudes above 305 m (1000 ft),
i.e., Mp < 5%. At lower altitudes Mp may in rease to 30% in order to obtain
higher tra king performan e.

P5- Roll angle response. In ase of engine failure in still air, the roll angle,
, should not ex eed 10 deg; its maximum steady state deviation should not
ex eed 5 deg. During engine failure, sideslip angle  should be minimised; the
steady state roll angle that is needed to a hieve this, should be redu ed to zero
with an overshoot of less than 50 % when the failed engine is restarted (the
failed engine's throttle setting steps ba k to that of the a tive engine). Under
moderate turbulen e onditions (see subse tion 14.2.6)  should remain smaller

168
than 5 deg.

P6- Airspeed response. The ontrolled system's airspeed, VA , should be able

to tra k speed ommands, VA , with a rise time tr < 12 s and settling time
ts < 45 s. There should be very little overshoot in the step response to speed
ommands at altitudes above 305 m (1000 ft), i.e., Mp < 5%. At lower altitudes
Mp may in rease to 30% in order to obtain higher tra king performan e. In
the presen e of a wind step with an amplitude of 13 m/s (25 kts) there should
be no deviation in the airspeed larger than 2.6 m/s (5 kts) for more than 15 s.
There should be no steady state error due to onstant wind disturban es.

P7- Heading rate. In ase of engine failure, the maximum heading rate, _,
should be less than 3 deg/se .

P8- Cross oupling between airspeed VA and altitude h. For a step in om-
manded altitude h of 30 m, the peak value of the transient of the absolute
error between VA and ommanded airspeed VA should be smaller than 0.5 m/s
(1 kt). Conversely, for a step in ommanded airspeed VA of 13 m/s (25 kts),
the peak value of the transient of the absolute error between h and h should
be smaller than 10 m.

Robustness riteria
R1- Centre of gravity variation. Stability and su ient performan e should
be maintained for horizontal entre of gravity variations between 15 and 31
% and verti al entre of gravity variations between 0 and 21 % of the mean
aerodynami hord (see table 14.5; we will not onsider variations in lateral
dire tion).

R2- Mass variations. Stability and su ient performan e should be main-
tained for air raft mass variations between 100000 to 150000 kg.

R3- Time delay. Stability and su ient performan e should be maintained for
transport delays from 50 to 100 ms.

Ride quality riteria

Ride quality riteria should ensure su ient passenger and pilot omfort. The
following spe i ations are designed to obtain an a eptable level.

Q1- Maximum verti al a eleration. Under normal onditions during manoeu-

vres (no turbulen e) the verti al a eleration at the entre of gravity should be
minimised; it should be less than  0.05 g1.
Q2- Maximum lateral a eleration. Under normal onditions during manoeu-
vres (no turbulen e) the lateral a eleration at the entre of gravity should be
minimised; it should be less than  0.02 g.
Q3- Damping. Unless stated dierently, there should be no overshoot in any
step response of any ontrolled variable at altitudes above 305 ft (1000 ft).

1
This value is used in industry during the design phase, in fa t the verti al and lateral
a eleration limits depend on frequen y. They are even lower at 2 Hz.

169
Below that altitude overshoot may in rease to 30 % in order to obtain higher
tra king performan e.

Safety riteria
S1- Airspeed. The airspeed must always be larger than 1:05  Vstall
, where
Vstall denotes the stall speed, i.e. the speed below whi h the air raft is unable
to maintain ight. This speed an be found from the following equilibrium
relation:
1 2 CLmax
mg = SVstall (14.58)
2
Substituting the relevant values from hapter 14.2, and assuming a mass of
120000 kg, we obtain Vstall = 51:8 m/s.
S2- Angle of atta k. In subse tion 14.2.3 it was given that the maximum lift
oe ient is obtained at an angle of atta k of 18 degree. Hen e, it an be
on luded from the previous requirement that the stall speed orresponds to:
stall = 18 deg. A value of 12 deg is onsidered a eptable.

S3- Roll angle. The maximum roll angle  should be limited to 30 deg.
S4- Sideslip angle response. At all times, sideslip angle  should be minimised.
For unit RMS intensity lateral Dryden gust the RMS of the sideslip angle in
losed loop should be less than that in open loop.

Control a tivity riteria

C1- A tuator eort minimisation. Under moderate turbulen e onditions (see
subse tion 14.2.6), mean a tuator rates for aileron, tailplane and rudder should
be less than 33 % of the maximum rates (see subse tion 14.2.5).

C2- Engine eort minimisation. Under moderate turbulen e onditions (see

subse tion 14.2.6), mean throttle rate should be less than 15 % of the maximum
rate (see subse tion 14.2.5).

14.3.3 Evaluation pro edure: RCAM mission and s e-

nario
To be able to evaluate all kinds of dierent ontrol design pro edures and
resulting ontrollers it is ne essary to nd a uniform evaluation pro edure, in-
dependent of the design method. An established pro edure to do this is to
dene a mission and a typi al landing approa h s enario (see [258, 37, 262).
This mission onsists of manoeuvres that an be evaluated by means of non-
linear simulations. The performan e of the ontrol law depends on the mission
phase, within whi h hard riteria or bounds on ertain signals should be met
and/or error signals must be minimised.

The mission and s enario to be own by the RCAM model onsists of a

landing approa h divided into the following segments (see gure 14.8)

170
 Trajectory for RCAM evaluation

1500

2 e
altitude (ZE) [m]

d
1000 1 f
3
c
b g h4 Runway
500 a
0
Wind
0
0
5 0
5
10 10
15 15
20
20 25
yposition (YE) [km] xposition (XE) [km]

Figure 14.8: the landing approa h for RCAM

 Segment I (0 to 1).
Starting at an altitude of 1000 m and with a tra k angle of  = 90 deg,
level ight is to be maintained with a onstant airspeed of 80 m/s . Dur-
1
ing this segment, the lateral features of the autopilot will be investigated
by simulating failure of the left engine (engine 1). This is indi ated in g-
ure 14.8: the failure o urs at point a, after whi h the engine is restarted
at point b. The transient and steady state behaviour of the system will
be analysed.

 Segment II (1 to 2).
This segment onsists of a ommanded o-ordinated turn from points
to d with a heading rate of _ =3 deg/se . The obje tives are to
maintain a onstant speed of 80 m/s, to keep the lateral a eleration
lose to zero, and to restri t the bank angle to  = 30 deg with onsistent
rudder/aileron dee tions.

 Segment III (2 to 3).

The des ent phase will be started a ording to the so- alled Frankfurt
des ent pro edure (see [35), whi h has been proposed for reasons of en-
vironmental noise redu tion. This des ent pro edure is engaged later and
is steeper than the lassi al des ent, whi h has a onstant glide slope an-
gle of = 3 deg. The starting altitude is h = 1000 m. After a short
period of level ight, the ight path angle is set to = 6 deg at point
e, and to = 3 deg at point f. The desired airspeed is 80 m/s.

1
The nominal airspeed during the landing phase depends on the air raft mass, it is taken
equal to 1.3 times Vstall : with a maximum landing weight of 150000 kg this results in  80
m/s.

171
  Segment IV (3 to 4).
The glide slope of = 3 deg is to be maintained during a wind shear
between points g and h. The air raft has to maintain safe ight and
should not deviate too far from the desired glide path. The wind shear
model used in the evaluation pro edure is a two dimensional model de-
rived from [201 (also see subse tion 14.3.4 for more information). The
desired airspeed is 80 m/s.

Throughout the evaluation pro edure a Dryden turbulen e eld, of s ale length
L = 305 m and amplitude  = 0:08 m/s, is assumed to be a tive. Note that
the amplitude is only 5% of the amplitude for moderate onditions as dened
in Chapter 14.2: this is done to prevent that the ee t of turbulen e on lateral
and longitudinal a elerations overrules other ee ts that we are interested in.
Superimposed on top of this turbulen e is a 10 m/s onstant wind with a xed
heading. This onstant wind is a tive in full respe t during Segments I and II
until point d, and is slowly redu ed to zero between points d of Segment II
and g of Segment IV (at the start of the wind shear model). The wind has no
verti al omponent and is dire ted along the negative earth-xed x-axis, i.e.,
it is a ross-wind during Segment I and a headwind during Segment III.
To he k robustness properties the entire approa h will be own with a most
forward, a nominal and a most aft horizontal entre of gravity lo ation. Fur-
thermore, one ight will be exe uted with a nominal entre of gravity lo ation
and a time delay of 100 ms.

14.3.4 Translation of design riteria into evaluation

riteria
It should be noted that it is not possible to he k all desired autopilot features
by ying a single landing approa h traje tory. Furthermore, the evaluation
pro edure should be relatively simple and straightforward: we want to be able
to apply it to a great variety of ontrollers. Hen e, the evaluation riteria
should be independent of the type of ontroller used: they should onsist of
al ulable indi ators that enable us to obtain an obje tive omparison between
ompletely dierent ontrollers.
For these evaluation riteria we will use the same lassi ation as was given
in the denition of the design riteria.

 performan e

 robustness

 ride quality

 safety

 ontrol a tivity

For ea h of these items and for ea h of the four traje tory segments a single
number will be al ulated. This number should not be onsidered to be the

172
nal word on overall autopilot performan e: it is merely an indi ator for one or
two important aspe ts. In most ases it is hosen su h that a value of smaller
than one is a eptable.
To further evaluate the dynami behaviour of the autopilot, we will onsider
several plots of key variables during ea h of the segments. We will ompare
the shape of the a tual traje tory with the demanded traje tory and provide
bounds that should be respe ted for good performan e. Similarly, we will plot
the most important deviations from the desired traje tory.

Segment I
For segment I we will plot a plan view of the referen e traje tory and the four
traje tories dened in subse tion 14.3.3, and then superimpose the bounds
given in gure 14.9. The points a and b orrespond to the beginning and end

First segment: top view

100

50
xdeviation [m]

0
0 a b 1

50

100

20 18 16 14 12 10 8 6 4 2
yposition (YE) [km]

Figure 14.9: Segment I: the ee t of engine failure with bounds

of the engine failure segment.

Performan e. The bound of 20 m is given to a ount for the ee t of turbulen e.

During engine failure, we allow a maximum lateral deviation of 100 m that
should be qui kly redu ed to less than 20 m at the end of the segment (when
the air raft should be stabilised again). With eyb denoting the lateral deviation
in body o-ordinates for the traje tory with nominal entre of gravity and time
delay we will use
 
max
j eyb (t)j jeyb (t1 )j
+ =2 (14.59)
t 100 20
as a measure that should be smaller than one for su ient performan e. Note
that during the entire segment the maximum deviation of 100 m should not be
ex eeded, and that at the end of the segment ( t1 orresponds with point 1) the
maximum deviation of 20 m is taken into a ount.

Performan e deviation. The maximum dieren es between the lateral deviation

of the traje tories with nominal and perturbed entre of gravity and nominal
and maximal time delays are onsidered:

eyb (t) := max (jeybmax(t) eyb (t)j; jeybmin (t) eyb (t)j) (14.60)

173
We will allow dieren es of 10% of the maximal allowable lateral deviations:
 

eyb (t)
eyb (t1 )
max + =2 (14.61)
t 10 2
should be smaller than one.

Ride quality. The maximum lateral a eleration will be limited by:

jny j < 1 (14.62)

0:2
i.e. jny j should be smaller than 0.2g: under normal ight onditions this value
should be mu h lower (0.02g, see subse tion 14.3.2), but engine failure is an
emergen y situation su h that an unusually large lateral a eleration is a ept-
able.

Safety. During the segment, the maximum angle of atta k  will be limited:

max

j(t)j 3 < 1 (14.63)
t 12
This implies we a ept  = 12 deg; the power is taken to stress the fa t that
 > 12 deg qui kly be omes una eptable (stall situation).

Control a tivity. The rudder a tuator eort will be onsidered that is needed
to stabilise the air raft after engine failure is lifted: this is al ulated as:

Z t1
R2 dt (14.64)
tb
with tb denoting the end of engine failure ( orresponds to point b in gure 14.9).
This value is not normalised to one as it is not lear what bounds an be
obtained: it will a t as a value for relative omparison of ontrollers.

Segment II
For segment II we will plot a plan view of the referen e traje tory and the four
traje tories dened in subse tion 14.3.3, and then superimpose the bounds
given in gure 14.10. Furthermore, to obtain a better insight in the results, we
will plot lateral deviations with bounds as given in gure 14.11.

Performan e. The maximum lateral deviation (due to the turn) and the lateral
deviation at the end of the segment (when the air raft should be stabilised
again) are onsidered:

 
max
j eyb (t)j jeyb (t2 )j
+ =2 < 1 (14.65)
t 200 20

174
 Second segment: top view Second segment: lateral deviations
1 300

0.5
d 2 200
0
yposition (YE) [km]

0.5

lateral deviation [m]

100
1

1.5 c 0
1 c d 2
2
100
2.5

3
1 200
3.5

4 300
24 23 22 21 20 19 18 17 16 0 1 2 3 4 5 6 7 8
xposition (XE) [km] along track distance from point 1 [km]

Figure 14.10: Segment II: plan view Figure 14.11: Segment II: lateral
of the 90 degree turn with bounds deviations during the 90 degree
turn with bounds

Note that during the entire segment a maximum deviation of 200 m should not
be ex eeded, and that at the end of the segment ( t2 orresponds with point 2)
a maximum deviation of 20 m is taken into a ount.

Performan e deviation. As in segment I, the maximum dieren es between

the lateral deviation of the traje tories with nominal and perturbed entre of
gravity lo ations and nominal and maximal time delays are onsidered. Again,
we will allow dieren es of 10% of the maximal allowable lateral deviations:
 

eyb (t)
eyb (t2 )
max + =2 < 1 (14.66)
t 20 2

Ride quality. As in segment I, the maximum lateral a eleration ny will be

onsidered:
jny j < 1 (14.67)
0:02
i.e. jny j should be smaller than 0.02 g.
Safety. As in segment I, the maximum angle of atta k during the segment is

j(t)j 3 < 1
limited:

max (14.68)
t 12

Control a tivity. The rudder and aileron a tuator eort is al ulated as:

Z t2


R2 + A2 dt (14.69)
t1
This value is not normalised to one as it is not lear what bounds an be
obtained: it will a t as a value for relative omparison of ontrollers.

175
 Segment III
For segment III we will plot a side view of the four traje tories dened in
subse tion 14.3.3. Figure 14.12 shows the referen e traje tory, the start and
end points of the segment (points 2 and 3) and the onsidered bounds; the
ommand a tions are labelled with e and f. We will also plot the verti al
deviation of the traje tories and overlay the bounds shown in gure 14.13.

Third segment: side view Third segment: altitude deviations

1100 30

2
1000 20
e

altitude deviation [m]

altitude (ZE) [m]

900 10 2 f 3

800 0

f e
700 10

600 3 20

500 30
17 16 15 14 13 12 11 10 16 15 14 13 12 11
xposition (XE) [km] xposition (XE) [km]

Figure 14.12: Segment III: side Figure 14.13: Segment III: verti-
view of the -6 and -3 degree glides- al deviations during the -6 and -
lope aptures with bounds 3 degree glideslope aptures with
bounds

Performan e. The maximum verti al deviation during the apture of the -6

degree glideslope and the verti al deviation at the end of the segment (when
the air raft should be stabilised again) are onsidered. Furthermore, speed
variations should be kept small in spite of the hange in required angle of
atta k. With ezb denoting the verti al deviation in body o-ordinates for the
traje tory with nominal entre of gravity and time delay, we will demand
 
max
j ezb (t)j jezb (t3 )j jVA VA j
+ + =3 < 1 (14.70)
t 20 6 4
for su ient performan e. Note that during the entire segment a maximum
deviation of 20 m should not be ex eeded, and that at the end of the segment
( t3 orresponds with point 3) a maximum deviation of 6 m is taken into a -
ount. Speed variations should not ex eed 4 m/s, i.e. 5% of VA = 80 m/s).

Performan e deviation. The maximum dieren es between the verti al devia-

tion of the traje tories for the nominal and perturbed entre of gravity lo ations
and nominal and maximal time delays are onsidered:

ezb (t) := max (jezbmax(t) ezb (t)j; jezbmin (t) ezb (t)j) (14.71)

We will allow dieren es of 10% of the maximal allowable verti al deviations:

 

ezb (t)
ezb (t3 )
max + =2 < 1 (14.72)
t 2 0:6

176
Ride quality. The maximum verti al a eleration nz will be limited:

jnz j < 1 (14.73)

0:1
i.e. jnz j should be smaller than 0.1 g.
Safety. Again, the maximum angle of atta k during the segment is limited:

max

j(t)j 3 < 1 (14.74)
t 12

Control a tivity. The tailplane a tuator eort is al ulated as:

Z t3
T2 dt (14.75)
t2
This value is not normalised to one as it is not lear what bounds an be
obtained: it will a t as a value for relative omparison of ontrollers.

Segment IV
For segment IV we will plot a side view of the four traje tories dened in
subse tion 14.3.3. The wind shear model, the desired traje tory through it,
and the bounds are given in gure 14.14. As mentioned before, the wind shear

Fourth segment: side view with windshear

1600

1400

1200
altitude (ZE) [m]

1000

800

600 3 g

400

200 h
4
0
11000100009000 8000 7000 6000 5000 4000 3000 2000 1000
xposition (XE) [m]

Figure 14.14: Segment IV: side view of the nal approa h with wind shear and
bounds

model is a two dimensional model derived from [201. Along the traje tory, the
air raft will be fa ed with a headwind going up to about W xE = 7 m/s, then
windspeed will hange to a tailwind of about W xE = 7 m/s, ombined with a

177
 downdraught of about W zE = 8 m/s (see gure 14.15). The result of this will
be a drasti de rease in air raft energy: the air raft will not be able to stay
on the desired traje tory. The size of the longitudinal deviation and the time
until re overy will be measures for evaluation of the ontroller.
For this reason we will also plot the longitudinal deviations with bounds as
given in gure 14.16.

Fourth segment: wind velocities during wind shear Fourth segment: altitude deviations
10 30
WXE
8 WZE
wind velocities WXE and WZE [m/s]

20
6

altitude deviation [m]

4
10 3
2
3 g h 4 g h 4
0 0

2
10
4

6
20
8

10 30
11000100009000 8000 7000 6000 5000 4000 3000 2000 1000 11 10 9 8 7 6 5 4 3 2 1
xposition (XE) [m] xposition (XE) [km]

Figure 14.15: Segment IV: wind Figure 14.16: Segment IV: verti-
speeds along the traje tory al deviations during the nal ap-
proa h with bounds

Performan e. The maximum longitudinal deviation (due to the wind shear)

and the longitudinal deviation at the end of the segment (when the air raft
should be within the de ision window) are onsidered for the traje tory with
nominal entre of gravity and time delay:
 
max
j ezb (t)j jezb (t4 )j
+ =2 < 1 (14.76)
t 20 1:5
Note that during the entire segment a maximum deviation of 20 m should not
be ex eeded, and that at the end of the segment ( t4 orresponds with point 4)
a maximum deviation of 1.5 m is taken into a ount.

Performan e deviation. As in segment III, the maximum dieren es between

the verti al deviation of the traje tories for the nominal and perturbed entre of
gravity lo ations and nominal and maximal time delays are onsidered. Again,
we will allow dieren es of 10% of the maximal allowable verti al deviations:
 

ezb (t)
ezb (t4 )
max + =2 < 1 (14.77)
t 2 0:15

Ride quality. As in segment III, the maximum verti al a eleration nz will be

limited:
jnz j < 1 (14.78)
0:2
i.e. jnz j should be smaller than 0.2g (in segment III this value is lower, but
wind shear is an emergen y situation).

178
Safety. We will onsider whether the air raft is within the de ision window at
the end of the segment. Lateral, verti al and speed variations are limited to 5
m, 1.5 m and 3 m/s respe tively as follows:

s  
1 eyb 2 ezb 2 (VA VA ) 2
( ) +( ) +( ) <1 (14.79)
3 5 1:5 3

Control a tivity. The tailplane and throttle a tuators eort is al ulated as

follows: Z t4


T2 + (T H1 + T H2 )2 dt (14.80)
t3
This value is not normalised to one as it is not lear what bounds an be
obtained: it will a t as a value for relative omparison of ontrollers.

179
15. The Classi al Control Approa h

Jim Gautrey 1

Abstra t. This do ument des ribes the design and analysis of a

lassi al ontroller developed for the Resear h Civil Air raft Model
(RCAM) design hallenge. The hallenge onsists of designing an
autopilot for an air raft whi h will enable it to meet pre-determined
requirements and to follow a pre-dened traje tory. Classi al on-
trol methods have been used for this parti ular design hallenge
entry, and were found to give satisfa tory performan e and robust-
ness.

15.1 Introdu tion

The design approa h adopted here for the RCAM design hallenge [145 is
based on lassi al ontrol methods. This method uses feedba k of measurable
variables plus ontrollers based on proportional, integral and derivative gains
to obtain the required performan e. The method is not inherently robust, and
therefore many dierent operating ases must be simulated with the ontroller
in order to verify its robustness. The design pro ess used for the RCAM on-
troller produ es an autopilot whi h is similar to an a tual air raft autopilot,
sin e the nature of the air raft leaves little s ope for variation in the design,
and lassi al ontrol methods are the most ommon design methods used for
autopilot development.
The ontrollers were designed using the Control Systems toolbox within
Matlab. No other toolboxes were required during the design pro ess.
Classi al ontrol methods are urrently used for the majority of ontrol sys-
tems designed for air raft as they were the only methods available for many
years, and therefore a large amount of knowledge and experien e has been a -
umulated. Their use within the air raft industry in ludes appli ations to the
design of major systems, su h as piloted ight ontrol, autopilot, and autosta-
bilisation su h as pit h and yaw dampers. These methods are also widely used
outside the aerospa e industry.
The main advantages of this design methodology are that it is well proven
and a large reservoir of knowledge exists about a tual implemented systems.
However, a prin ipal disadvantage is that a reasonably large amount of prior
knowledge is required on erning the operational and behavioural hara ter-
isti s of the plant sin e the design pro ess is arried out intera tively by the

1
Air Vehi le Te hnology Group, Craneld College of Aeronauti s, Craneld University,
UK.

180
 ontrol system designer. It is the designer who has to spe ify the hanges in the
ontroller during the design pro ess in order to generate the required perfor-
man e and robustness, as opposed to methods where the ontroller is generated
and optimised automati ally from a series of onstraints determined by the
designer.

15.1.1 Controller Stru ture

Classi al ontrol methods usually result in an autopilot ontroller of the form
seen in gure 15.1. It onsists of inner loops, whi h generally augment the
stability of the air raft and outer loops, whi h generally ontrol the air raft
ight path. The system model may, or may not in lude gain s heduling or other
non-linearities su h as a tuator rate and position limiting. Gain s heduling
was not required in the ontroller developed here. However, rate and position
limits were required, and the development of these elements is des ribed. Gain
s heduling would be required in an autopilot intended for operation over a large
ight envelope due to the variation in the aerodynami hara teristi s of the
air raft.

FEEDBACK LOOPS

INNER LOOPS AIRCRAFT

OUTER LOOPS OUTPUTS

COMMANDS

Figure 15.1: Classi al autopilot ontroller stru ture

S heduling an be organised in two prin ipal ways for an autopilot. The rst
is gain s heduling when the autopilot is in one parti ular mode of operation,
su h as height hold. In this ase, individual ontroller parameters will be varied
to take a ount of the non-linear hara teristi s within the air raft, su h as
variations in its hara teristi s with speed or onguration hange. The se ond
form of s heduling is mode swit hing. For example, if the autopilot hanges
from height hold to glideslope hold, a dierent set of ontrol laws may be used.
In pra ti e, this is a omplished through the sele tion of an alternate outer
loop. The design presented in this do ument has limited gain s heduling to
take a ount of autopilot mode hanges, but it does not take a ount of gross
speed or air raft re onguration hanges. These forms of gain s heduling are
not restri ted solely to air raft.

15.1.2 Complex ontroller onsiderations

The ontroller designed here are simple ontrollers. However, there are many
real world onsiderations whi h need to be made. This se tion highlights the
areas where the design of a simple ontroller su h as the one illustrated above
needs to be modied to ope with real life situations.

181
Gain s heduling
Classi al ontrollers, at their simplest level, are designed at a single operating
point. The RCAM lassi al ontroller is no ex eption to this. In order to a -
ount for hanges in operating onditions, gain s heduling is used to modify
the ontroller gains as the operating point of the plant hanges. This is used
sin e it is usually a straightforward pro ess as there are relatively few gains
in a lassi al ontroller, and the gain s hedules an be very straightforward.
For example, most air raft autopilots require a gain s hedule whi h a ounts
for variations in the air raft airspeed, and this is often enough to ensure good
performan e over a wide range of operating onditions. However, there is some-
times a requirement to s hedule gains with respe t to other parameters, and
this an sometimes ompli ate the ontroller design.

Pra ti al onsiderations
A tual ontrol systems suer from problems whi h perfe t systems do not.
These in lude non-perfe t dynami s of the plant under onsideration whi h
an result in the ontroller having to ontrol a plant whi h is dierent to the
one whi h it was originally designed for. Also, outputs from sensors are rarely
perfe t, and therefore this needs to be taken a ount of. Filtering may be
required to ompensate for noise in the signals.
In addition, the lassi al ontroller onsidered for the RCAM problem does
not take a ount of airframe stru tural modes, whi h it may ex ite. Again,
ltering is used to take a ount of this by removing outputs to the air raft
ontrol surfa es at the frequen ies whi h would ex ite the stru tural modes.

Outer loop variations

For ontrollers su h as an autopilot, it is desirable to have several dierent
modes, or to design the ontroller so it may perform several tasks. This is
usually a omplished through the use of dierent outer loops. The same inner
loop is used for the dierent modes, but dierent outer loops are engaged
depending on the task requirements.
This is a ommon pra ti e amongst lassi al ontrol designers as it enables
the ontroller stru ture to be kept simple while enabling spe ied tasks to be
performed well. It has been used in the lassi al ontroller for the RCAM
design.

15.2 The Sele tion of the Controller Ar hite -

ture for the RCAM Problem
The autopilot has been split into a longitudinal and a lateral ontroller, with
no intera tion between the two. Ea h is des ribed individually. Both the
longitudinal and lateral ontrollers have inner and outer loops.

182
 Available measurement signals have been used, see referen e [145. For an
a tual implementation, the design would use the most suitable of the available
signals as some are better than others, i.e. have less noise or are easier to
measure. For the purposes of the design hallenge, all output signals have
been assumed to be perfe t. The only de ien y in the signals available is the
la k of a lateral a eleration and pit h attitude signal. These are ommonly
available to autopilots sin e they are relatively easily measured, and are usually
onsidered to be essential. The pit h attitude signal has been synthesised
for the purposes of the design hallenge, but attempts to synthesise lateral
a eleration were unsu essful.
Inner and outer loops have been used in the ontroller design pro ess. The
inner loops are used to provide the ne essary stability augmentation, while the
outer loops regulate the augmented air raft's ight path performan e. Fun -
tional blo k diagrams for the longitudinal and lateral ontrollers an be seen
in gures 15.2 and 15.3.

Height error Controller for step height demands

K + K +
Height rate
- S
Attitude
Limiter
K +
+
Tailplane
K +
Demand
-
Pitch Attitude
Height error
P+I+D Mode
Controller Switch K +
Pitch Rate
Controller for ramp height demands

Rate
Limiter

+
Airspeed demand
Throttle Demand
+ P+I+D
Airspeed Controller

Figure 15.2: Longitudinal ontroller blo k diagram

15.3 The Translation of RCAM Design Criteria

into Method Dependent Obje tives
The RCAM design riteria are set up in method independent terms in sub-
se tion 3.2 of the RCAM manual [145. This subse tion des ribes how these
method independent riteria may be translated into method dependent require-
ments.

183
Track error
K + Roll Attitude
K + Limiter
- S
Track rate +
K + Aileron
K + Demand
-
Roll Attitude

K +
Roll Rate
P+I
Controller
Sideslip Angle K +
Washout
Filter Rudder
+ Demand
Roll Angle K
Washout
Filter
Yaw Rate K +

Figure 15.3: Lateral ontroller blo k diagram

Classi al ontrol methods do not rely on the denition of weighting matri es

or other optimisation routines sin e the optimisation pro ess is arried out by
hand, and therefore there is no dire t translation of design riteria into the
design pro ess in this respe t. However, as the ontroller design progresses,
the various riteria are veried, and ontroller design iteration is based on its
performan e with respe t to those riteria. This is addressed in the following
subse tions where ea h of the dierent riteria are dis ussed. Most riteria rely
on the examination of step responses.

Performan e Criteria
The ontroller stru ture is partly determined by the performan e requirements.
For example, transient and steady state requirements often di tate whether an
integrator should be used in a parti ular ontrol loop. For a step height de-
mand, an integrator is not stri tly required. However, in order to a hieve zero
steady state error following a ramp height demand (su h as a glideslope), at
least one free integrator is required in the ontroller. Therefore this an gener-
ate oni ting ontroller design requirements for dierent autopilot fun tions,
whi h require a solution su h as outer loop swit hing or gain s heduling.
After the ontroller stru ture has been determined, requirements su h as
overshoot and response limits an be ontrolled dire tly by the gain sele tion
pro ess. This is one of the easiest riteria to design for sin e responses, su h
as the air raft height response to a parti ular demand an easily be evaluated

184
from the earliest stages of lassi al design using the te hniques presented here,
and therefore the ee t of gain hanges an easily be assessed. For example, if
a parti ular response is ex essively sluggish, it may indi ate that a parti ular
value for a damping gain within a ontroller is too large.

Robustness Criteria
Classi al ontrol te hniques are not inherently robust. Therefore it is ne essary
to test the designed ontroller at many dierent ight ases to ensure that the
design is suitable for ea h ase. For this parti ular design exer ise, a mid-
envelope ight ase is used for ontroller development, and then the result is
he ked at a series of ight ases throughout the ight envelope (see table 15.1),
orresponding to a wide variation in air raft mass, entre of gravity position
and omputational time delay.
The stability of the ontroller an be he ked at ea h stage by he king the
pole lo ations on the s-plane of the linear air raft model plus autopilot. The
bandwidth of ea h omponent of the autopilot plus air raft an also be veried.
In addition, gain and phase margins may be used.

Passenger Comfort Criteria

The passenger omfort riteria are primarily governed by the design of the
inner loops sin e these are the loops whi h regulate the air raft's manoeuvre
performan e and its transient response to atmospheri disturban es. Therefore
it is ne essary to he k for passenger omfort as the design progresses as with
the performan e riteria, and to iterate until the design meets the requirements.
The outer loops do ae t the passenger omfort riteria, but with the orre t
inner loops, the outer loops should not have a large ee t on it. Limiters may
also be required in order to limit demand su h as pit h rate pla ed on the
ontroller in order to meet the omfort requirements.

Safety Criteria
The safety riteria primarily govern the air raft attitude and speed and are
primarily ae ted by the design of the inner loops. For example, it is normal
to limit the air raft attitude demands whi h the outer loops an make on the
inner loops. This does not permit the outer loop to demand an ex essive pit h
attitude when the air raft may not have su ient thrust to hold the required
airspeed. For the lateral tra k loops, this onsists of limiting the maximum roll
demand that the ontroller an make. Therefore by areful design of the inner
loops, and sele tion of appropriate limiters between the outer and inner loops,
the air raft safety requirements an be attained.
The air raft airspeed an be ontrolled pre isely by the autothrottle de-
sign. By designing the autothrottle loop well, the airspeed an be pre isely
ontrolled for normal ight ases. For abnormal onditions, su h as windshear
or mi roburst, simulation will demonstrate whether the air raft meets the re-
quirements. A orre t inner loop design will ensure that the air raft should

185
meet the requirements for normal ight. For abnormal ight, su h as an engine
failure, it is ne essary to simulate and therefore demonstrate that the air raft
meets the requirements. However, due to the intuitive nature of lassi al on-
trol te hniques, analysis of the air raft response an be used to highlight any
problems, and often a solution an be found from this.

Power Consumption Criteria

The power onsumption riteria are not signi antly ae ted by the ontroller
design. If there is a problem with the design, su h as a large os illatory motions
or transients, then the higher power onsumption may be the result of ex essive
ontrol a tivity on one parti ular ontrol surfa e. For example, this would
indi ate a problem within the inner ontrol loops, su h as ex essive gain, and
may therefore be modied by adjusting the inner loop gains a ordingly.

15.4 The Design Cy le Des ription

This subse tion des ribes the a tual design y le. Therefore the design y le
for ea h ontroller omponent is onsidered, along its appropriate parameter
adjustment strategy.
For ea h air raft model omponent, the inner loops are designed rst using
linear analysis, primarily using robustness, omfort and some performan e ri-
teria. The design is then tested with the full non-linear air raft simulation to
ensure that the design works in this domain. Safety onsiderations are onsid-
ered in order to design any limits whi h need to be pla ed on the inner loop
demands. The outer loops are then designed in the linear domain, primarily
with performan e requirements in mind. Finally, the resulting ontroller is
transferred to the simulation environment and formally evaluated with respe t
to all of the design riteria.
For ea h of the longitudinal and lateral ontroller designs, the following
steps are performed.

1. Assemble the appropriate linear model from the non-linear air raft model.
This gives a model in state spa e form whi h an be used to represent
the air raft.

2. Augment the air raft model with the required sensor and a tuator dy-
nami s, and a linearised time delay if appropriate.

3. Design the inner stabilising loops using the te hniques des ribed below.
These are then assessed against the appropriate riteria.

4. Design the outer autopilot loops against the inner stabilising loops, and
re he k that the model meets the requirements.

5. Che k the ontroller for robustness by simulating at a number of ight

ases over the ight envelope and with dierent air raft ongurations.

186
 6. Perform the nal evaluation.

Iteration is required during this pro ess, both within and between steps. A
suitable air raft ight ase must also be used for the baseline design pro ess.
The ase used has an air raft mass of 120000kg, a g position at [0.23,0,0 (see
[145), VAIR = 80 m/s and h = 1000 m. This is in the middle of the ight
envelope, and seems to provide a reasonable design point.

15.4.1 Longitudinal Controller Design

Before des ribing the detailed design of the longitudinal ontroller, it is ne es-
sary to des ribe the prin iples behind its overall operation. It onsists of three
main omponents. The rst omponent is the pit h inner loop, whi h augments
the longitudinal air raft stability, and whi h has the same ee t as a pit h atti-
tude demand ontroller. The se ond omponent is the autothrottle, whi h uses
airspeed feedba k to throttles to regulate the airspeed. The third omponent is
the pit h outer loop, whi h regulates the air raft height by translating height
error into a pit h attitude demand.
In designing the longitudinal ontroller, it is ne essary to design the inner
loops rst. These an then be used to augment the air raft, and the outer loops
therefore designed around this augmented air raft. Experien e has shown that
it was not ne essary to onsider the time delay during the initial design, but
slight gain modi ations were made during the simulation of the inner loops
with dierent ight ases to allow for time delay variations.

Pit h Attitude Inner loop

The pit h attitude inner loop omprises two feedba k paths. The rst is pit h
rate feedba k, whi h is used to augment the damping of the short period mode,
and the se ond is the a tual pit h attitude demand loop. A tuator or time delay
dynami s were not required sin e it was found that their ex lusion from the
linearised model used for ontroller design did not have a deleterious ee t
on the ontroller performan e. A short period damping of lose to unity is
desirable sin e this removes any residual pit h os illations from the  response,
and therefore the gain required an be found from the interse tion of the short
period root lo us with the real axis of the plot. In addition, high phugoid
damping is required.
Testing the inner loop ontroller in the simulation environment showed that
improved pit h attitude tra king ould be obtained by modifying the gains pre-
viously derived. In reasing the gains gave in reased  tra king performan e.
This modi ation pro ess was done iteratively with the autothrottle opera-
tional. The autothrottle design is des ribed in the next subse tion.

Speed Inner Loop

The speed ontrol omponent is the se ond omponent of the longitudinal au-
topilot. Autothrottles have traditionally in luded feedba ks to the throttles

187
from airspeed, and sometimes the pit h attitude error and height error. How-
ever, this autothrottle onsists of feedba k of the speed error to the throttles
through a Proportional + Integral + Derivative ontroller. This ensures good
speed tra king sin e the integrator will ensure that the speed error is redu ed
to zero while the dierentiator will smooth the transient response to a speed
hange demand, i.e. ensuring that the longitudinal a eleration is regulated.
Modi ations were made to the baseline autothrottle design on the results from
non-linear simulation to step speed and height responses, and also from simu-
lating a range of dierent ight ases.

For the design onsidered here, with a maximum permitted longitudinal

a eleration of  0.02g, a rate limiter pla ed in the demanded speed signal path
having limits of 0:1962 ms 2 satisfa torily limits the longitudinal a eleration
to a eptable limits.

Height Outer Loop

The third omponent of the longitudinal ontrol system is the height ontrol
outer loop. For the outer loop design, two ases were onsidered initially. The
rst is the response to a step demand in height. In order to a ommodate
air raft trim hanges, it is ne essary to in lude an integrator with the system,
otherwise, the air raft may have a onstant steady state error in height if a
speed hange is made. This is di ult to do without getting an overshoot in
height. However, the ontroller stru ture used in gure 15.2 does give zero
overshoot with zero steady state error for trim hanges in a height hold situa-
tion.

The se ond ase was the air raft response to a ramp height demand, or zero
steady state error to a non-zero ight path angle demand. This required the
addition of an integrator in the height error loop whi h integrates the height
error signal to ensure zero steady state error.

In order to limit the demanded attitude, an attitude limiter was pla ed

between the height outer loop and inner pit h attitude loop. This was limited
to approximately +30 degrees (nose up), -12 degrees (nose down). The limits
were derived from the autopilot authority, i.e. with the limits in pla e the
air raft ould hold the sele ted pit h attitude without signi antly a elerating
or de elerating. In pra ti e, lower limits ould be used sin e su h large attitudes
are rarely required when under autopilot ontrol.

However, analysis showed that the normal a eleration aused by the on-
troller is ex essive, and therefore a rate limiter is pla ed between the inner
pit h loop and the outer height loops. This limits the peak normal a eleration
by limiting the pit h rate. It must be remembered that the ontroller outer
loop design must not make ex essive demands on the inner loop sin e when the
limiter is a tive, the air raft is, in open-loop ight, and in a sense, un on-
trolled. However the inner loops, retain full authority ontrol of the air raft at
all times.

188
15.4.2 Lateral Controller Design
As with the longitudinal ontroller, the lateral ontroller has three omponents,
shown in the on eptual design, gure 15.3. The rst omponent is the yaw
damper. This in ludes the use of the rudder for turn o-ordination, sideslip and
lateral a eleration ontrol. The se ond omponent is the roll attitude inner
loop, whi h is used to regulate the roll attitude through the use of aileron. The
nal omponent is the lateral outer loop whi h is used to regulate the lateral
tra k, and feeds into the roll attitude inner loop.

Yaw Damper and Turn Co-ordination Controller

A yaw damper has one main purpose, to in rease the Dut h roll damping in
order to enable the autopilot and pilot to y the air raft. A se ond fun tion is
to support turn o-ordination, whi h is often alled yaw autostabilisation. For
the purposes of this report, the term yaw autostabilisation will refer to both
fun tions. There are several dierent methods whi h may be used to design
the yaw autostabiliser. The nal ontroller used sideslip angle feedba k to the
rudder to a hieve this sin e it serves as a useful way to minimise the sideslip
angle, whi h then a hieves turn o-ordination. The use of this feedba k has the
ee t of destabilising the aperiodi spiral motion, and in reasing the frequen y
of the Dut h roll mode, whi h in this ase gives an in rease in dire tional
stability.
The design an be arried out in the following steps.

1. Sideslip angle to rudder feedba k gain sele tion.

The beta to rudder root lo us plot was used to sele t the required Dut h
roll natural frequen y. Sele ting a value for the Dut h roll natural fre-
quen y of about 2 rad/s gives a value of 13 rad/rad for the  to rudder
gain.

2. Yaw rate to rudder gain sele tion.

The next stage is to sele t the values for the yaw rate to rudder gain.
This has the ee t of in reasing the Dut h roll damping ratio. A desirable
value of Dut h roll damping is approximately 0.7. It is also ne essary to
in orporate a high pass lter sin e the  gain previously sele ted will be
the only gain whi h is not ltered. Filtering is required sin e the other
signals have a onstant value in a steady turn whi h must not be passed
through to the rudder. Therefore a washout lter with a time onstant
of 0.2 s is used. The hoi e of time onstant was arbitrary, and this is
typi al of a value used in pra ti e. The yaw rate to rudder root lo us
plot was used to do this.

3.  feedba k to the rudder.

The addition of washed out  feedba k is the next stage. This has the
ee t of modifying the lateral a eleration hara teristi s of the ontroller.

4. Addition of  integrator.
The next stage is to add an integrator into the beta feedba k to rudder

189
 loop. This is in luded to meet the engine failure requirements sin e this
ase requires a large amount of rudder to be held in order to redu e the
ee ts of the failure. An I/P ratio of 0.2 was found to be suitable.

5. Veri ation with simulation.

The a tuator ould then be added in, and the ontroller operation veried
in the simulation environment. Analysis showed that it had little ee t
on the time responses. However the  to rudder gain was in reased to
1.3 rad/rad in an attempt to optimise the lateral a eleration.

It has been stated that the gains used here in rease the RMS  response
ompared to the unaugmented air raft, whi h is not desirable. However,
this was ne essary due to the la k of lateral a eleration signal.

Roll Attitude Inner Loop

In the same way as the pit h attitude is ontrolled with pit h attitude and pit h
rate feedba k to the tailplane, the air raft is ontrolled in roll by using roll
attitude and roll rate feedba k to the aileron. Ordinarily, feeding roll attitude
error to aileron (with a suitable gain) results in an air raft whi h is stable in
roll attitude sin e roll rate is essentially proportional to aileron input (for a
onventional air raft up to moderate roll angles). However, the RCAM air raft
is not onventional laterally. The roll angle response to a small aileron step
input is a step hange in roll angle, due to the spiral mode whi h is rapid and
stable. Therefore it is ne essary to augment the hara teristi s of the air raft
before the roll attitude autopilot an be designed.
The stages in designing the inner roll loop are as follows. It has been
assumed that the yaw autostabilisation has been arried out.

1. Roll rate feedba k to aileron.

This provides a more step-like roll rate response to a step aileron demand.

2. Roll attitude feedba k to ailerons.

Then a  demand loop is losed around the augmented ailerons so that
the input an be used to demand a given roll angle.

3. A tuator Ee ts.

The a tuator was then introdu ed to the design. It had little ee t on
the step response. This on ludes the design of the roll inner loops in the
linear domain.

Simulation of a roll angle demand showed that the simulation was very
lose to the linear response, and the design was suitable. However, iteration
was performed in order to qui ken the response as with the pit h attitude inner
loop. This on luded the design of the inner roll loop.

190
Lateral Tra k Hold Autopilot
The outer lateral loop ontrols the air raft's lateral tra k and turn performan e.
It is a lassi al tra k ontrol loop, where lateral tra k error, with lead ompen-
sation, is fed into the roll attitude demand loop. This gives a zero steady state
tra k error when the air raft is tra king a straight line, sin e any tra k error
auses the air raft to a elerate towards the demanded tra k, and the damping
prevents the air raft from overshooting the demanded tra k ex essively. This
type of lateral tra k holding has been ommonly used in the past.
The lateral tra k error is onverted into a roll angle demand by multiplying
it by a predetermined gain. However, in order to be able to ope with a non-
symmetri al ight ase, su h as engine failure, it is ne essary to in lude an
integrator, as with the height step ontroller. A damped omponent of the
lateral tra k error was also in luded in the tra k error to  demand sin e trials
showed that the response tra k error was very under-damped without it.

Heading Demand Outer Loop

A separate heading demand outer loop was designed for the ontroller. This
then generates a bank angle demand signal whi h is then fed into the inner roll
loop to ontrol the air raft heading. The s hemati for the heading outer loop
is similar to the height step demand outer loop.

15.5 Linear Analysis

For the purposes of linear analysis, the ight ases from table 15.1 were on-
sidered. All are at an airspeed of 80 m/s.

Flight Case Mass (Kg) CGx CGy CGz Time Delay (ms)
e1 120000 0.23 0 0 0.05
e2 120000 0.23 0 0 0.1
e3 120000 0.31 0 0 0.05
e4 120000 0.15 0 0 0.05
t1 100000 0.23 0 0 0.05
t2 150000 0.23 0 0 0.05
t3 100000 0.15 0.03 0.21 0.1
t4 100000 0.15 0 0 0.05
t5 150000 0.15 0.03 0.21 0.1
t6 100000 0.31 0 0.21 0.1
t7 100000 0.31 0 0.21 0.05
t8 150000 0.31 0.03 0.21 0.1
t9 150000 0.31 0.006 0.21 0.1

Table 15.1: Flight ases onsidered for analysis

191
 A series of evaluation runs were performed. The longitudinal analysis on-
sidered height steps, airspeed steps and glideslope inter epts. The design re-
quirements listed in [145 were all met. The lateral analysis onsidered heading
angle and lateral tra k steps, and again all of the requirements listed in the
design manual were met.

15.6 Analysis of the Resulting Controller in

terms of the Applied Methodology
During the design phase, the prin ipal analysis te hniques are omprised of the
following methods. The inner loop designs were, on the whole, arried out in
the frequen y domain using root lo us te hniques to modify the air raft eigen-
value hara teristi s. Then the design emphasis swit hed to the time domain,
sin e this is the domain in whi h many of the performan e riteria are dened,
and the ee t of hanging ontroller gains an be more readily understood.
However, the eigenvalues were observed from time to time during this pro ess
to ensure that they remained stable. Problems with ontroller performan e,
su h as overshoot in a parti ular response an generally be ontrolled by in-
tuitive gain modi ation. The problem hara teristi and knowledge of the
air raft provide the information to assess whi h gain needs adjustment.
Simulation was used extensively during the design phase to fa ilitate the
observation of the ee ts of gain variation. Linear simulations were used ini-
tially with step and ramp input fun tions. Non-linear simulations were used
subsequently in order to ne tune the gains and to provide a nal he k that
the ontroller performan e in the non-linear domain was a eptable.
Ride quality was assessed using the simulation with various inputs, su h as
a step heading demand and a step height demand. A eptable ride quality an
be a hieved by sensible gain sele tion and by the orre t hoi e of feedba k
variables. Again, simulation was used to show whether a parti ular hange
made to the ontroller had an ee t on ride quality. Analyti al methods are
presented in [166, but they were found to require modifying on the basis of
simulation results.
Control a tivity an be analysed by monitoring during a parti ular manoeu-
vre. In the ase when there is high ontrol a tivity on one parti ular surfa e,
that a tivity an be redu ed by better gain sele tion, or in extreme ases, by
the use of a lter su h as a low pass lter. Resort to su h xes was not re-
quired during the design phase for the RCAM ontroller, sin e, in ases where
there was ex essive ontrol a tivity, it ould generally be redu ed by a better
gain sele tion in the troublesome feedba k path.

15.6.1 Satisfa tion of Design Obje tives

Performan e
The primary emphasis is on performan e, spe i ally path tra king, and all
of the requirements were met in the nal evaluation. However, there were

192
 ompromises with some of the other riteria, spe i ally ride omfort, and to
a lesser extent, robustness. The main robustness variation was in transient
performan e and the steady state performan e was generally onstant.

All of the lateral deviation riteria spe ied in the design manual were met.

Altitude Response. All of the altitude riteria spe ied in the design manual
were met.

Heading angle. The rise time and settling time riteria were met for all the
ight ases onsidered.

The open-loop RMS heading angle with unit RMS lateral turbulen e inten-
sity (W20 = 8 m/s) was found to be 0.50 degrees. With the heading angle hold
engaged, it was found to be 0.89 degrees. Therefore the ontroller does not
meet the turbulen e riterion. This was found to be due primarily to the inte-
grator in the sideslip to rudder path. If this integrator was removed, the RMS
error redu es to 0.54 degrees, whi h just ex eeds the non-augmented ondition.

Flight Path Angle Response. The overshoot riterion was met for all ight ases.
However the rise time riterion was not met. This riterion was relaxed in order
to allow for the ride quality riterion to be met for the glideslope a quire, as it
is required to limit normal a eleration to less than  0.1g. Therefore, the rise
time was 8 se onds instead of the stipulated 5 se onds.

Roll angle with engine failure. The maximum roll angle divergen e and steady
state value was met with the ontroller following engine failure. However the
maximum overshoot of 50% is not met with the lateral tra k loop sele ted in the
autopilot sin e the air raft has to roll ba k to attain zero tra k error, ausing
an overshoot in roll attitude. If a onstant heading is demanded then the roll
attitude overshoot riterion is met (overshoot in roll is 40%) sin e there is no
orre tion required to attain zero lateral tra k error.

Airspeed response. The rise time riterion was met for all CG ases. However,
the settling time riterion was not met for the non-zero lateral CG positions.
This was unexpe ted sin e the integrator in the speed ontrol loop should for e
the error to zero. Again, the trim routine was suspe ted. The wind step
deviation riterion was met. From the evaluation routine, it was observed that
the largest airspeed deviation during the nal evaluation, in luding the wind
shear, was 0.8 m/s.

Heading rate response. The heading rate riterion under engine failure was met
sin e the maximum heading rate with an engine failure was 0.5 deg/s. This is
observed during the evaluation phase.

Cross oupling from airspeed to altitude. The requirements to minimise the

airspeed error for a given altitude step input and to minimise the attitude
error for a given airspeed step were met.

193
Safety
The safety requirements were not designed for spe i ally. However, during
the nal evaluation, it was veried that the requirements were met.

Airspeed Regulation. The airspeed regulation riterion was always met as the
airspeed was always within  0.8 m/s of the demanded airspeed.
Angle of atta k. This riterion was met in the nal evaluation. The angle of
atta k never ex eeded 3.5 degrees or be ame less than 0.5 degrees.

Roll angle. The requirement to limit the roll angle to less than 30 degrees was
met.

Sideslip angle. The sideslip angle is ontinually minimised by the a tion of the
proportional gain and integrator following engine failure. The RMS value of
the lateral gust velo ity was al ulated and a W20 value [145 of 8 m/s gave
a value of 1.02m/s for the RMS lateral gust intensity, and therefore this value
was used to al ulate the response. The RMS value of  was found to be 0.58
degrees for the open-loop air raft, and 0.89 degrees for the augmented air raft.
Therefore the requirement was not met. The trade-o here is made in the yaw
autostabiliser design between lateral path tra king and the need to redu e the
engine failure transient, see subse tion 15.4.2.

Robustness
The prin ipal design emphasis was pla ed on satisfying the robustness require-
ments during the design pro ess. After the performan e requirements had been
satised, only minor adjustments were made to the ontroller design in order
to meet the robustness requirements. There was some ompromise between
robustness, omfort and performan e, and a ontroller whi h seemed to give
the best overall balan e was sele ted.

Time Delay. The requirement to be able to ope with a time delay of up to 0.1
se onds were met. This aused some problems with the height step response in
the nal evaluation, and onsequently the ontroller was modied by in reasing
the gains in the height outer loop, whi h gave a more onsistent response for
dierent air raft masses and CG positions.

Mass Variations. The ability to ope with variations in air raft mass are met.

Centre of Gravity Variation. The ability to ope with variations in the entre
of gravity positions were met for CGx and CGz variations. For variations in
the lateral entre of gravity position, (CGy), problems were en ountered whi h
were thought to be due to the trim routine.

Ride Quality Criteria

As the design progressed, the omfort requirements were he ked. However, it
was found that the lateral a eleration requirement ould not be met due to the
unavailability of lateral a eleration signal for feedba k purposes. This would

194
be available in a onventional autopilot, and experimentation showed that the
lateral a eleration ould be redu ed if this signal were available. Some of the
peak values of normal a eleration ex eeded the limit, but not ex essively so.

Longitudinal A eleration. A maximum longitudinal a eleration of 0.02g was

a hieved, ex ept during the turn, where there was a (-0.08g peak), and during
the windshear, where there was a +0.1g / -0.04g peak.

Verti al a eleration. The riterion was met, ex ept during the turn when
the verti al a eleration ex eeded the permitted limits. The required load
fa tor to maintain the turn rate is approximately 0.1g, and during this phase,
the maximum verti al a eleration is less than 0.2g. The normal a eleration
during the third segment was also slightly greater than the permitted value.

Lateral A eleration. This was less than 0.025g for the majority of the nal
evaluation. It peaked at approximately 0.055g during turn entry and exit, but
was redu ed to about 0.025g as the turn be ame established. In order to redu e
it further it would be ne essary to in orporate lateral a eleration feedba k into
the lateral ontroller. The lateral a eleration requirement during engine failure
was met.

Damping. This riterion was met for the majority of the ight ases. See the
performan e riteria subse tion for more detail.

Control A tivity During the Final Evaluation

Control a tivity was within the permitted levels at all times, and it was mostly
well under the maximum permitted values.

Control A tivity in Turbulen e

Problems arose during ontroller performan e evaluation in turbulent ondi-
tions.Having a high  to rudder gain resulted in higher rudder displa ements
for a given  gust, whi h oered no improvement to the RMS  gust response.
The integrator is in luded in the ontroller in order to omply with the engine
failure requirement. The high gain is in the ontroller to redu e  in order to
redu e the lateral a eleration and also to ope with the engine failure ase.
The rudder and aileron rates were measured in the simulated moderate
turbulen e, and were less than the maximum values.

15.7 Results of the Automated Evaluation Pro-

edure
This se tion presents the methodology independent results of the ontroller
designed in the previous se tions. It is mostly based on the evaluation mission
and s enario dened in [145: both overall tra king performan e and inner-
loop behaviour of the ontrolled system will be evaluated by means of bounds
on key variables.

195
Segment I: the ee t of engine failure.
As the RCAM air raft model is twin-engined, a single engine failure will mainly
result in lateral deviation. Hen e gure 15.4 provides a top view of the rst
traje tory segment.

First segment: top view Second segment: lateral deviations

300

100
200

50

lateral deviation [m]

100
xdeviation [m]

0 0
0 a b 1
1 c d 2

100
50

200
100

300
20 18 16 14 12 10 8 6 4 2 0 1 2 3 4 5 6 7 8
yposition (YE) [km] along track distance from point 1 [km]

Figure 15.4: Left: segment I - the ee t of engine failure. Right: segment II -
lateral deviations during the 3o /s turn

Third segment: altitude deviations Fourth segment: altitude deviations

30 30

20 20
altitude deviation [m]

3
altitude deviation [m]

10 2 f 3 10
4
g h
0 0

10 e 10

20 20

30 30
16 15 14 13 12 11 11 10 9 8 7 6 5 4 3 2 1
xposition (XE) [km] xposition (XE) [km]

Figure 15.5: Left: segment III - verti al deviations from the desired glideslope.
Right: segment IV - verti al deviations from the desired glideslope.

Segment II: the 3 deg/s turn.

Figure 15.4 provides a loser look at the a tual lateral deviations during the
se ond segment of the ight, whi h is a 90 deg turn at 3 deg/s.

Segment III: the apture of the -6 and -3 degrees glideslope.

The verti al deviations from the desired glideslope are plotted in gure 15.5,
whi h is the third ight segment.

Segment IV: the nal approa h with windshear.

While on nal approa h with a glideslope of -3 deg the ee t of a windshear
model is onsidered. The verti al deviations from the desired glideslope are

196
plotted in gure 15.5.

Numeri al results
Here a table of numeri al results based on the dis ussed simulation results is
given 15.2. For the motivation and al ulation prin iple of the various results
see [145.

Segm. I Segm. II Segm. III Segm. IV Total

Performan e 0.0907 0.1847 0.1523 0.1228 0.1376
Perf. Dev. 0.0727 0.0381 0.1608 0.0933 0.0912
Comfort 0.3732 1.2615 1.1771 0.5771 0.8472
Safety 0.0041 0.0301 0.0063 0.0419 0.0206
Power 0.0027 0.0061 0.0148 0.0298 0.0133

Table 15.2: Numeri al results of the evaluation pro edure

15.8 Con lusions

Classi al Control methods are the most ommon methods used for autopilot
design. They are not inherently robust, but the ontrollers that are produ ed
an be made robust with orre t ontroller stru ture, time and eort. In addi-
tion, they are not initially easy to use, sin e mu h knowledge is required about
both the methods and the air raft in question. However, one this knowledge
has been obtained, for one parti ular air raft, the methods may readily be ex-
tended to other air raft or ight ases without too mu h redesign due to the
simple nature of the ontrollers produ ed, and the high visibility of the design.
These methods are also rigorously analysable, and therefore they an be
readily ertied, and sin e they ontain relatively few omponents, the ee ts
of failure of some of those omponents an be assessed relatively easily. There
is a great deal of experien e on erning their use and implementation available
within most vendors and airframe manufa turers.
The ontroller designed here was at a single ight ase. The pro ess to
extend this design to a number of ight ases is essentially simple due to the
relatively small number of gains required in the ontroller, and in addition,
there are simple ways of s heduling the gains for variations with airspeed.
Their prin ipal disadvantage is the time taken to perform the design pro-
ess, although if an outline design is available this time an be redu ed. It is
ommon in industry for an existing autopilot design to be modied to suit a
new air raft, as opposed to a ompletely new design being performed, and this
redu es the design time. A signi ant amount of knowledge on erning air raft
and their hara teristi s is also required to support the design pro edure sin e
the optimisation of the ontroller depends on the knowledge and intuition of
the designer and not a omputer algorithm.

197
 The ontroller designed here meets the majority of the requirements spe i-
ed for the RCAM hallenge, and any deviation has been noted. The majority
of the dis repan ies would not pose insurmountable problems in pra ti e sin e
the way in whi h the ontroller would be operated is dierent. For example, a
glideslope inter ept would generally be arried out from below the glideslope,
and the autopilot would know in advan e what is being demanded of it. Also,
the autopilot would not have to ope with an engine failure in the manner
required here sin e a tual engine parameters would be available to it.
Sin e an a tual autopilot would not have to ope with these requirements
in quite the same way, the design ould be tailored more towards the other
requirements su h as ride omfort. Problems were experien ed with ertain ride
omfort levels, and these would be alleviated if other signals had been made
available to the ontroller. Other improvements to the autopilot presented here
ould also be made through improved gain sele tion, for example.
Finally, this ontroller demonstrates that a simple, robust ontroller an be
produ ed with lassi al ontrol design te hniques, although there is a trade-o
between some design and performan e requirements.

198
16. Multi-Obje tive Parameter Synthesis
(MOPS)

Hans-Dieter Joos 1

Abstra t. Multi-obje tive parameter synthesis (MOPS) provides

a systemati way for omputational ontrol law tuning by dire tly
spe ifying and iterating bounds and demands on stability quality,
ontrol performan e, and physi al ontrol realization onstraints.
The design an be based on linear or non-linear models and the on-
troller stru ture an be hosen in a onvenient way taking advantage
of the design engineer's knowledge and prior expertise. Given de-
sign spe i ations an be transformed dire tly into the mathemati-
al design riteria needed for performing the method. Variations in
stru tured parameter sets and operating onditions are taken are
of by a multi-model formulation. The multi-obje tive/multi-model
parameter tuning is done in a goal-oriented way by applying param-
eter optimisation. In this ontribution it is shown that the MOPS
systemati an be used to develop robust ight ontrol systems with
good performan e and passenger omfort.

16.1 Introdu tion

The RCAM ben hmark problem addresses the design of an autopilot for the -
nal approa h of a transport-type air raft. The designed ontroller is required to
be robust with respe t to variations in speed, weight, entre of gravity position
(both verti al and horizontal), time delays, non-linearities, and engine failure.
The method whi h is applied is an optimisation-based multi-obje tive/multi-
model/multi-parameter design methodology alled Multi-Obje tive Parameter
Synthesis (MOPS). Be ause of the inherent freedom in spe ifying ontroller
stru tures, this approa h an be used in ombination with any ontroller syn-
thesis method, and be ause of the possibilities in spe ifying the design goals in
a onvenient multi- riteria manner this yields a systemati goal-oriented way
to solve the ontrol design tuning problem.
Two dierent approa hes are adopted for synthesising ontroller parameters.
For the lateral ontroller, dire t optimization of the parameters of an appropri-
ate ontroller stru ture was applied. For the longitudinal ontroller an indire t

1
DLR German Aerospa e Resear h Establishment, Institute for Roboti s and System
Dynami s, Control Design Engineering (Prof. G. Grbel), D-82234 Wessling
E-mail: dieter.joosdlr.de

199
approa h was hosen: An analyti synthesis method (we used LQR-synthesis)
denes the ontroller stru ture and its parameters. The free parameters of
the synthesis method i.e. the entries of the weighting matri es Q and R serve
as tuners that are optimized a ording to a set of design riteria. Instead of
the LQR synthesis method, other methods like parameterized eigenstru ture
synthesis ould have been used as well.
The ontrollers have been designed by taking into a ount both performan e
and robustness expli itly. The performan e spe i ations given in the RCAM
Design Challenge Manual are dire tly used as multiple design obje tives to be
treated simultaneously. Robustness is treated by using a multi-model approa h.
That is, a ommon ontroller is designed simultaneously for a hosen set of
evaluation models. These models hara terize the variations in weight and
entre of gravity.
In multi-obje tive design ea h one of the design-obje tives has to be math-
emati ally des ribed as a single riterion. For the present design task, the
spe i ations for air raft performan e, safety, and ontrol eort as given in
Chapter 14 an be easily transformed to suited mathemati al riteria fun tions
whi h serve to ompute an individual gure of merit for ea h design obje -
tive. Expli itly taking are of ea h design obje tive by a orresponding gure
of merit makes the ne essary design- ompromise de isions very transparent.

16.2 Applied Control Design Methodology

16.2.1 Controller Stru ture and Parameterisation
Both linear and nonlinear ontrollers (e.g. fuzzy-logi ontrollers [129) an be
used. If a spe i analyti al ontrol law synthesis te hnique is applied within
this framework, the ontroller stru ture is naturally bound to this synthesis
te hnique.
Here ontrol laws are parameterised in two dierent ways. In a lassi al
P-I-D ontrol stru ture with additional shaping lters the P-I-D gains and the
lter parameters are tuned to satisfy losed-loop performan e and robustness
requirements. This is for lateral ontrol. An LQR output feedba k ontrol law
for longitudinal ontrol is parametrized by the state- and ontrol eort weights
Q and R.
The free design parameters (i.e. the P-I-D gains and the LQR-weights) are
omputed by a min-max parameter optimisation set up.

16.2.2 System Model Des ription and Robustness

Both linear and nonlinear design models are used. There is no methodi re-
stri tion on the representation of system disturban es.
Robustness against stru tured parameter variations or, e.g., sensor fail-
ures is a hieved by applying a ommon ontroller to a set of worst- ase xed-
parameter models. Here, this set represents a parameter-nominal model and

200
two dynami worst- ase models, i.e. the slowest and the fastest one, deter-
mined by a nonlinear a priori parameter study. For ea h model an appropriate
set of riteria an be spe ied individually. Thereby the multi-model problem
is transformed to a multi-obje tive problem.
In general, there exists no theory that guarantees stability or performan e
robustness over the range of operation, if only a nite number of operating
points is onsidered simultaneously. It depends on the physi al properties of
the system to be designed whether runaways an exist. If they exist, they
have to be added to the set of operating points treated by the multi-model
approa h. In this design study we dete ted no runaways.
Robustness of the ontroller around an operating point an be enfor ed
in the multi-obje tive approa h by applying suitable robustness riteria su h
as gain/phase margins in addition to the hosen performan e riteria.

16.2.3 Design Spe i ation

The design spe i ations an be requirements in state spa e (e.g. eigenstru -
ture), time-domain (e.g. step response, ontrol rate), and frequen y-domain
(e.g. bandwidth, stability margins et ). Ea h design obje tive is to be mathe-
mati ally des ribed by a positive denite riterion i to be minimized or to be
onstrained by an upper bound, .f. Se tion 16.4.

16.2.4 Design Cy le Des ription

The designer formulates a MOPS setup by dire tly spe ifying the design goals
as a set of positive-dened omputational riteria. By this multi-obje tive
formulation all the various oni ting design goals are taken are of individually.
To ea h riterion i a demand value di is asso iated by the designer. Then the
tuning parameters T of the ontrol law are omputed by solving the min-max
parameter optimization problem

min max f i (T )=di g

T i

subje t to performan e and tuning restraints:

gj (T )  dj ; Tkmin  Tk  Tkmax:
The solution provides a best-possible pareto-optimal ontrol parameter
tuning. It provides dire t quantitative information about the design oni ts
and onstraints for a hosen ontroller stru ture. Iterating the demand values
di in an evolutionary manner for es the trade-o solution into a desired goal
dire tion. For a more detailed introdu tion, see Chapter 2.
A typi al design loop is shown in Figure 16.1 illustrating that the MOPS
methodology omplements hosen modelling-, ontrol synthesis-, and analysis
methods by losing the loop via a goal oriented ontrol parameter tuning.

201
 synthesis plant
model model

d T P controller closed-loop M simulation/

MOPS synthesis
model model analysis

c performance/cost
I
criteria

Figure 16.1: MOPS loses the design loop by ontroller parameter tuning.

16.2.5 Pra ti al Software Aspe ts

Appli ation of MOPS requires the set-up of a omputation loop a ording to
Figure 16.1 and availability of a suitable min-max parameter optimization soft-
ware. The overall omputing time mainly depends on the riteria evaluations.
Hen e fast algorithms and software implementations are to be used for the
orresponding analysis omputations.
In our ase, synthesis, analysis, simulation, and optimization, are performed
in ANDECS
2 [99 whi h provides appropriate design-data management fa ili-
ties. In this environment, omputations an also be performed by MATLAB
3
or by MATRIXX
4. The non-linear air raft model, the a tuators, and the
parametrized ontrol stru ture are modelled in Dymola . From Dymola models
5
DSblo k- ode [186 is generated automati ally. This simulation-model ode is
dynami ally linked to ANDECS or MATLAB/SIMULINK. The latter is used
to perform the automated evaluation pro edure des ribed in Se tion 16.7.

16.3 Controller Ar hite ture

The ontroller is split into a longitudinal and a lateral ontroller. For onve-
nien e and visibility in design and maintenan e, these ontrollers are treated
separately from ea h other.
Sin e MOPS does not pres ribe a spe i ontroller stru ture, we de ided
to apply two dierent approa hes. This also shows the exibility of the design
method regarding ontroller ar hite tures.

16.3.1 Lateral ontroller

For the lateral ontroller we hoose onventional stru tures with P-I elements,
wash-out lters, or just onstant gains as they are do umented in textbooks.
See for example [165 and [35. The parameters of these elements are used

2
ANDECS is a registered trademark of DLR.
3
MATLAB, SIMULINK are trademarks of The MathWorks, In .
4
MATRIX X is a trademark of Integrated Systems, In .
5
Dymola is a produ t of Dynasim AB, Lund, SE

202
as tuning parameters during optimisation. The ontroller is omposed of an
inner loop ontroller for stabilising and handling-qualities augmentation, and
an outer loop ontroller for guiding the air raft along a given traje tory. The
stru ture of the overall lateral ontroller is shown in Figure 16.2. The gain of

psidotc

2 2 da
ylat 80 /v
phic
outer loop inner loop
2 2 dr
80 /v
ylatc

p phi r beta

Figure 16.2: Stru ture of the lateral ontroller of dynami order 7.

the ontrol ommands is s heduled by the fun tion 802=VA2 to over hanges in
the ontrol surfa e ee tivity due to speed variations. VA denotes the measured
airspeed, and 80 m/s is the airspeed the ontroller is designed for. The overall
lateral ontrolleris 7th order in luding two integrators and a prelter for lateral
tra king ommands.
The inner loop, see Figure 16.4, onsists of a bank angle integral ontrol
system with onstant gain feedba k roll rate damper and a sideslip suppression
system with yaw damper and wash-out network to allow hanges in yaw rate
when hanging the air raft's heading. A ross oupling gain between bank
angle and rudder is introdu ed to improve the de oupling of both motions.
The outer loop ontrols the lateral deviation from the ommanded path.
For this a P-I-D ontroller and a 2nd order lter is introdu ed. The ne essity
of the lter for stabilisation has been found by root lo i onsiderations. Only
the lateral deviation is used as feedba k variable, see Figure 16.4.
The ontroller distinguishes lateral ommands (no hange of heading) and
turn ommands. For lateral ommands, a rst order prelter is used. For turns
also the demanded heading rate is ommanded.

16.3.2 Longitudinal ontroller

The stru ture of the ontroller results from the synthesis method hosen: LQR
state feedba k with a hosen synthesis model, see Figure 16.5.
The synthesis model is the linearized 5th order longitudinal model in lud-
ing the altitude z of the air raft. The parameters of the synthesis model are
hosen right in the middle of the given parameter ranges: m = 125000 kg,

x = 0:23;
z = 0:105. A design airspeed of 80 m/s is assumed. This 5th
order synthesis model is augmented by linear rst order lag lters to model the
a tuators with time onstants given in Chapter 14. It is further augmented by
integrators for airspeed and altitude ontrol, respe tively. The resulting on-

203
 p 2 2 da
kr + 80 /V

phic +
kr1 1/s
-

phi
kr2 +

kyr

r dr
ky s 2 2
+ 80 /V
s + ay

beta
ky1

Figure 16.3: Stru ture of the inner lateral ontroller

psidotc

atan(v/g * psidotc)

turn off

s 2+fz1s+fz0 phic
1 +
+ fd
Tvf s+1 2
ylatc s +fn1s+fn0

ylat

Figure 16.4: Stru ture of the outer lateral ontroller

commands
2 2 dT
. 80 /v
zc x = Ax+Bu+Ec
- dTH1
y = Cx+Du+Fc actuators
vac
- dTH2

q
- nx measurements
- nz
- wv
-

z0 va0 q0 nx0 nz0 wv0

trim conditions

Figure 16.5: Stru ture of the longitudinal ontroller of dynami order 4.

204
troller is therefore of 4th order, i.e. 2 integrators and an internal model of the
a tuators. There is no need for a seperate design of inner and outer loops.
The LQR-synthesis method yields a state feedba k ontrol law. However,
employing the provided 6 measurements q; nx ; ny ; wv ; z; and VA ; it is possible
to substitute the state by the available measurements using a (pseudo) inverse
of the measurement matrix. The overall longitudinal ontroller is therefore
given by the following equations

2 1 0 0 0
3 2 1 0
3
TT TT
6 0 1 0 07 6 0 1 7

T

x_ = 6 TT H 7x+6 TT H 7
4 0 0 0 0 5 4 0 0 5 T H
0 0 0 0 0 0
2 3
0 0  
0 077 VAC VA
6
+ 6
4 1 05 z z
0 1
 
T
= Ky1
(ylon ylon
trim ) + K x
y2
T H
+
Ky = [Ky1 ; Ky2 = Kx
Clon
where
+ denotes
Kx is given by the Ri ati solution of the LQR problem, and Clon
the pseudo inverse of the linearized longitudinal measurement matrix Clon of
the synthesis model.
The tuning parameters for optimization are now the entries of the weighting
matri esQ and R. We de ided to use an output weighting of the form
T Q Clon together with diagonal Q and R matri es.
[Clon
The longitudinal ontroller distinguishes altitude ontrol and glide slope
ontrol. For altitude ontrol z is the onstant altitude for whi h the air raft
is trimmed. For glide slope ontrol, z is the urrent altitude of the demanded
glide slope. Gain s heduling for airspeed adaptation, similar to the lateral ase,
is used for tailplane ontrol.

16.4 Translation of Design Obje tives to Math-

emati al Synthesis Criteria
MOPS design is well suited to deal with various design obje tives simultane-
ously. Ea h design obje tive is to be mathemati ally des ribed by a well-dened
riterion i with a value whi h is the smaller, the better the obje tive is a om-
plished. For the RCAM ben hmark problem it is straightforward to translate
the spe i ations given in Chapter 14, Se tion 14.3.2, into mathemati al ob-
je tive fun tions that an be used for optimisation.
For the longitudinal ontroller 18 types of riteria have been used simulta-
neously, .f. Table 16.1. In this table, the se ond olumn des ribes the spe i-
 ations as they are given in the RCAM Manual. The third olumn gives the

205
mathemati al formulation of the orresponding riteria. The 4th olumn shows
the used demand values given as bounds in the Manual; if no spe i value is
given min indi ates that the riterion value should be minimized. For min
riteria an appropriate demand value has to be hoosen in an evolutionary
manner during design. The last olumn shows the identi ation names given
to ea h riterion. The eigenvalue riteria No. 17, 18, are auxiliary riteria.
They are used to prevent unstable solutions and to enfor e good damping.
Overall, there are 18 riteria types to judge the longitudinal behaviour of
the ontrolled air raft. However, all obje tive fun tions involved are standard
riteria in the multi-obje tive design environment of ANDECS_MOPS [99 and
thus an be used o-the-shelf  without extra formulation eort.
For a multi-model approa h this set of riteria is applied to ea h one of the
models involved. Hen e in our design set up with 3 models we end up with 54
riteria to be taken are of simultanously.
For the lateral ontroller we used similar riteria types to rate the step
responses of bank angle  and lateral deviation dy, the disturban e reje tion to
wind, stability and damping of eigenvalues, ontroller eort, or omfort. The
omfort riterion is expressed by means of the lateral load fa tor as

Z tend
= n2y (t)dt :
0
It should be mentioned that also frequen y-domain riteria or any nonlinear
spe i ation ould be dealt with in the multi-obje tive design. However, for
the RCAM ben hmark problem only time domain spe i ations are given.

16.5 Des ription of the Design Cy le

16.5.1 Pro edure
A priori, the RCAM ontroller ar hite ture was split into a longitudinal and
a lateral ontroller to be designed independently. For both ontrollers the
following design de isions have to be made:

(i) Sele tion of an appropriate linear or nonlinear ontroller stru ture. If

an analyti synthesis method is used, an appropriate synthesis model for
synthesising the ontroller parameters has to be hosen also.

(ii) Sele tion of a set of dynami worst- ase evaluation models; via these
evaluation models the riteria fun tions are evaluated. They should be
su ient to ree t the properties of the original (nonlinear model) and
yet simple to enable fast omputation during optimisation. The set of
evaluation models should represent the possible parameter un ertainties
to a su ient amount.

(iii) Sele tion of the design riteria. The mathemati al design riteria have to
be hosen su h that the given design spe i ations are met orre tly.

206
No. Spe i ations Mathemati al Criteria Demand ID
1 altitude unit step: R tend
zero steady state error, settling = t (h(t) 1)2 dt min HH1

t0 = 10s, tend = 30s

1
time < 45 s

2 altitude unit step:

rise time < 12 s = t2 t1 12 RHH1
h(t1 ) = 0:1, h(t2 ) = 0:9
3 ross oupling altitude air-
speed: for a step in om-
manded altitude of 30 m, the = max jVA (t)j 0.5/30 HV1
peak value of the transient
t
of the absolute error between
VA and ommanded airspeed
should be smaller than 0.5 m/s
4 airspeed unit step: R tend
zero steady state error, settling = t (VA (t) 1)2 dt min VV1

t0 = 10s, tend = 30s

0
time < 45 s

5 airspeed unit step: = t2 t1 , VA (t1 ) = 0:1,

rise time < 12 s VA (t2 ) = 0:9 12 RVV1
6 ross oupling airspeed alti-
tude: for a step in om-
manded airspeed of 13 m/s, the = max jh(t)j 10/13 VH1
peak value of the transient of
t
the absolute error between h
and ommanded h should be
smaller than 10 m
7 altitude unit step:
overshoot < 5% = max h(t) 1.05 OHH1
t
8 airspeed unit step:
overshoot < 5% = max VA (t) 1.05 OVV1
t
9 airspeed wind disturban e:
for a wind step with amplitude = max VA (t) 2.6 VW1
of 13 m/s there should be no
t>15
deviation in the airspeed larger
than 2.6 m/s for more than 15
s
R tend
10 altitude wind disturban e: = 0 h2 (t)dt min HW1
no expli it spe i ation given tend = 30s
ontrol a tivity riteria:
eort minimization for
11 tailplane, altitude ommand R tend min DEZ1
12 throttle, altitude ommand = 0 u2 (t)dt min THZ1
13 tailplane, airspeed ommand R tend min DEV1
14 throttle, airspeed ommand = 0 u_ 2 (t)dt min THV1
15 throttle, wind step min THW1
16 throttle rate, wind step min DTHW1

 
Re(evi )
17 relative stability of eigenvalues
evi : =1 min 0.6 DAMP1
no expli it spe i ation i jevi j
18 absolute stability of eigenva-
lues evi : = exp(max(Re(evi ))) 0.95 BOUND1
no expli it spe i ation
i

Table 16.1: Design riteria used for longitudinal ontroller

207
 (iv) Sele tion of the demand values. For ea h design riterion a demand value
has to be asso iated. By iterating the demand values, the resulting om-
promise trade-o solution an be for ed into a desired dire tion. The
demand value an be

- an upper bound to a design riterion (method of inequalities),

- a weighting value that expresses the designer's preferen es related

to the other riteria to be minimised.

16.5.2 Performing the design y le for longitudinal on-

trol
Sele tion of the evaluation models
The evaluation models are used to al ulate the losed loop design-obje tives
during optimisation. The set of models should be representative for the param-
eter variations assumed. In order to determine a set of representative linear
longitudinal models we assessed, for an airspeed of 80 m/s, the un ontrolled
air raft pit h rate step responses and eigenvalue lo ations for a su ient num-
ber of parameter variations in mass and entre of gravity. This proved that
the family of time responses and eigenvalues are overed by only two param-
eter onstellations. Taking those together with the hosen nominal parameter
onstellation yields a 3-model problem with parameters

mass
x
z
nominal 1 125 000 0.23 0.105
variation 2 100 000 0.31 0
variation 3 150 000 0.15 0.21

The air raft models are augmented by linear a tuator models and a 2
nd
order Pad-approximation to des ribe a time delay  = 100 ms.
In multi-obje tive design, evaluation models need not ne essarily be lin-
ear. However, numeri al simulation of linear systems an be done more e-
iently than non-linear simulation and is therefore to be prefered in optimisa-
tion. Moreover, for linear models one an ompute eigenvalues whi h serve as
additional stability indi ators.

Sele tion of design riteria

The design riteria des ribed in Se tion 16.4 are sele ted. These basi riteria,
amongst others, are available in the design environment ANDECS_MOPS.
They ree t the design spe i ations in a su iently dire t manner.

Sele tion of the demand values and design iterations

The demand values are now asso iated to the riteria values. If the spe i-
ations in Chapter 14 provide quantitative values, those are used as demand
values. For example, in the presen e of a wind step with an amplitude of 13 m/s

208
there should be no deviation in the airspeed larger than 2.6 m/s for more than
15 s. This demand results immediately in the riterion

= max jV (t) VAC j  2:6

t>15 A
with demand value 2.6.
Other riteria su h as riteria for the ontroller eort that are not spe ied
by a spe i value are weighted relative to ea h other. Su h demand values are
iterated in an evolutionary manner during the following design steps in order
to for e the design into a desired dire tion.
Several onse utive design iterations have been done. This means several
optimisation runs with dierent demand values were arried out, ea h one
resulting in a trade-o solution a ording to the hosen demand values. The
general design pro edure is illustrated by an example design step:
The design goal is to improve the rise time of airspeed from about 12 s to
less than 8 s. To a hieve this, only the demand values orresponding to the
rise-time riteria ( riterion RVV1 of table 16.1) have to be hanged from 12 to
8. Starting an optimisation run yields - after about 3 min of omputing time -
the result indi ated by the thi k lines in Figure 16.6. The airspeed response is

thc wind step thc wind step thc wind step

E E E
thc

thc

thc
0.00 0.00 0.00
-0.30 -0.30 -0.30
-0.60 -0.60 -0.60
0.50 1.25 2.00 E1 0.50 1.25 2.00 E1 0.50 1.25 2.00 E1
t t t
wind step wind step wind step

E1 E1 E1

0.75 0.75 0.75

v,z

v,z

v,z

0.25 0.25 0.25

-0.25 -0.25 -0.25
0.50 1.25 2.00 E1 0.50 1.25 2.00 E1 0.50 1.25 2.00 E1
t t t
velocity step velocity step velocity step

E E E

0.8 0.8 0.8

0.4 0.4 0.4
0.0 0.0 0.0
0.50 1.25 2.00 E1 0.50 1.25 2.00 E1 0.50 1.25 2.00 E1
t/s t/s t/s
height step height step height step

E E E

0.8 0.8 0.8

0.4 0.4 0.4
0.0 0.0 0.0
0.50 1.25 2.00 E1 0.50 1.25 2.00 E1 0.50 1.25 2.00 E1
t/s t/s t/s
BOUND1

BOUND2

BOUND3
IDEXP

DTHW1
DAMP1

DTHW2
DAMP2

DTHW3
DAMP3
RHH1

RVV1

OHH1
OVV1

DEZ1
THZ1
DEV1
THV1
THW1

RHH2

RVV2

OHH2
OVV2

DEZ2
THZ2
DEV2
THV2
THW2

RHH3

RVV3

OHH3
OVV3

DEZ3
THZ3
DEV3
THV3
THW3
HH1

HV1
VV1

VH1

VW1
HW1

HH2

HV2
VV2

VH2

VW2
HW2

HH3

HV3
VV3

VH3

VW3
HW3

Figure 16.6: Comparing dierent design out omes by time response indi ators.
The diagrams in one olumn belong to the same evaluation model.

mu h faster, rise time is less than 8 s. Altitude response and reje tion of wind
disturban e are almost un hanged. However, the ontrol eort is in reased for

209
both tailplane and throttle a tivity between 10% and 25%. This is indi ated
by the normalized riteria values represented in the parallel oordinate display
also in luded in Figure 16.6.

16.5.3 Performing the design y le for lateral ontrol

Sele tion of evaluation models
Similar onsiderations as for longitudinal ontrol lead to a 3-model problem
with parameters

mass
y
nominal 1 125 000 0
variation 2 150 000 -0.03
variation 3 100 000 0.03

For evaluation, the linearized air raft models are augmented by rst order
lag lters for modelling the a tuators for aileron and rudder. Contrary to the
evaluation models for the longitudinal ontroller no additional time delay was
taken into a ount during design.

Sele tion of the design riteria

For the inner roll and yaw rate damper no expli it spe i ations are given in
Chapter 14. The riteria we sele ted are eigenvalue riteria to improve damping
and absolute stability margin. For bank angle ontrol, step-response riteria are
sele ted as introdu ed in Se tion 16.4. The same holds for sideslip suppression
and lateral deviation. Control eort was taken into a ount by integral-square
riteria.

Sele tion of the demand values and design iterations

The demand values are asso iated with the orresponding riteria value by
applying the same methodology as in the ase of longitudinal ontrol.
A ording to the ontroller stru ture sele ted, the design of the lateral on-
troller was divided into three major design steps.
In a rst step, the roll and yaw rate damper is designed su h that a relative
damping of 0.6 and a minimum absolute stability margin of 0.25 was a hieved
for all evaluation models simultaneously.
In a se ond step, yaw suppression and bank angle ontrol is designed. Us-
ing the standard riteria des ribed above, it is straightforward to a hieve the
a ording spe i ations.
The third step deals with the design of the lateral tra king performan e.
The parameters of the inner loop, designed before, remain xed during these
design iterations.

210
16.5.4 Design of a turn ompensation using the nonlinear
air raft model
Nonlinear assessment of the ontrollers designed so far for a trimmed 3 deg/s
turn showed poor damping for lateral or altitude step responses. To improve
damping the ontroller stru ture is augmented by a turn ompensator feeding
bank angle  on tailplane dee tion T as shown in Figure 16.7.

trim condition phi0

phi kuko1s+kuko
dT
2 2
s+kuko2 phi0* 80 /v

Figure 16.7: Stu ture of turn ompensator.

Sele tion of evaluation models

To tune the parameters of the ompensator we use the nonlinear model of the
ontrolled air raft within the MOPS design framework. For this the Dymola
model was augmented by the ompensator. The air raft was trimmed for a 3
deg/s turn for nominal parameter values and nominal speed. Using the non-
linear air raft model was the simplest way of treating the oupled lateral and
longitudinal motions during a turn. Furthermore it illustrates the exibility of
the MOPS method on erning appli ation of models and ontroller stru tures.

Sele tion of the design riteria

Design riteria steady-state errors of lateral and altitude step responses are
omputed as integral riteria similar to riterion No. 1 of Table 16.1.

Sele tion of the demand values and design iterations

All demand values are set to the urrent riteria values. This means that
all riteria are normalized and treated with the same preferen e. This is a
ommonly used pro edure in MOPS when no spe i demand values are invoked
by the spe i ations.

After a few design iterations, an optimized parameter set was found that
improves damping onsiderably. This is shown by the lateral step response in
Figure 16.8.

211
 Lateral deviation step response Deviation from altitude during lateral step

E1 E1

DY

dz
1.00 0.75

0.75 0.50

0.50 0.25

0.25 0.00

0.00 -0.25

-0.25 -0.50

-0.50 -0.75

-0.75 -1.00
0.0 0.4 0.8 1.2 1.6 2.0 2.4 2.8 3.2 3.6 4.0 E 1 0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 E 1
TIME TIME

Figure 16.8: Deviation of y and z for a lateral step response during a trimmed
3 deg/s turn with and without turn ompensation.

16.6 Analysis of the Controlled Air raft in

Terms of the Applied Methodology
16.6.1 On-line Analysis
In multi-obje tive ontroller design the obje tive fun tions are omputed from
indi ators obtained by system analysis and simulation. The resulting anal-
ysis data are therefore available, even during optimization. This is utilized
in ANDECS_MOPS for monitoring the optimization pro ess on-line. Figure
16.6 shows su h an on-line visualization with two optimization results. Whi h
analysis data should be plotted and how they should be arranged, is under
ompletely free ontrol by the user.
To ea h indi ator orresponds a set of design riteria. By dire tly ompar-
ing riteria values with the indi ators plotted in the diagrams, one an easily
he k whether the obje tive fun tions are well dened. By means of su h a vi-
sualization of the underlying dynami s indi ators one an also gain knowledge
about the proper size of the demand values whi h are not expli itely given in
the spe i ation. If, for example, steady-state behaviour is good enough by
inspe tion, the demand value for the steady-state riterion an be xed to the
resulting riterion value in the subsequent design iterations.

16.6.2 Parallel Coordinates

To visualize the whole set of riteria values over a omplete optimization run,
we use parallel oordinates [122, [79. The oordinate axis of the n-dimensional
riteria spa e are arranged in parallel. A point in the riteria spa e, i.e. the
riteria values for a given set of tuning parameters, is represented in parallel
oordinates by a line whi h inter onne ts all individual riteria values. Parallel
oordinates are well suited to elu idate

 how well riteria are fullled: the deeper the line the better the result.

 where oni ts between riteria exist: the lines ross.

212
  where on urrent riteria exist: the lines are (almost) parallel.

Figure 16.6 shows the parallel oordinate representation of all riteria in-
volved in the longitudinal design. The thi k oordinate line orresponds to
thi k time responses in the diagrams above.

16.6.3 Nonlinear Analysis

In order to dete t nonlinear ee ts, an a ompanying assessment based on
the non-linear RCAM model was performed for some important requirements.
Figure 16.9 shows 32 automated omputation experiments for minimum and
maximum parameter values of mass,
x,
z ,
y, as well as minimum and
maximum time delay for xed design airspeed of 80 m/s. The various diagrams

1.1 Altitude step [30m] 1.2 Altitude for VA step [13m/s] 1.3 Deviation from 3deg/s turn

E3 E3 E2
Z

1.020 1.002 0.25

deviation [m]
1.012 0.996 -0.25

1.004 0.990 -0.75

0 1.0 2.0 3.0 4.0 5.0 E1 0 1.0 2.0 3.0 4.0 5.0 E1 0 0.2 0.4 0.6 0.8 1.0 E2
TIME TIME t [s]

2.1 Speed step [13m/s] 2.2 VA for altitude step [30m] 2.3 VA for wind step uw [13m/s]

E2 E1 E1
V

V
0.90 8.010 7.75

0.86 7.980 7.25

0.82 7.950 6.75

0 1.0 2.0 3.0 4.0 5.0 E1 0 1.0 2.0 3.0 4.0 5.0 E1 0 1.0 2.0 3.0 4.0 5.0 E1
TIME TIME TIME

3.1 Y step [10m] 3.2 Psi step [0.1 rad] 3.3 Gamma step [0.1 rad]

E1 E -1 E -1
Y

PSI

GAMMA

0.75 0.75 0.75

0.45 0.45 0.45

0.15 0.15 0.15

0 1.0 2.0 3.0 4.0 5.0 E1 0 1.0 2.0 3.0 4.0 5.0 E1 0 1.0 2.0 3.0 4.0 5.0 E1
TIME TIME TIME

4.1 Engine failure: phi, beta 4.2 Engine failure: Y deviation 4.3 Engine failure: heading rate

E -1 E2 E -1
0.4 0.8 0.30
Y

psidot

-0.4 0.4 0.00

-1.2 0.0 -0.30

-2.0 -0.4 -0.60

0.0 0.4 0.8 1.2 1.6 E2 0.0 0.4 0.8 1.2 1.6 E2 0.0 0.4 0.8 1.2 1.6 E2
TIME TIME Time
phi beta

Figure 16.9: Nonlinear assessment for 32 parameter variations and dierent

ight manoeuvres, in luding engine failure.

show the following analysis results with dotted lines indi ating the allowed
toleran es:
1.1 altitude response for a step ommand of 30 m
1.2 altitude response for a airspeed ommand of 13 m/s
1.3 lateral deviation in ase of a 3 deg/s turn
2.2 airspeed response for an altitude step of 30 m
2.3 airspeed response for a headwind step of 13 m/s
3.1 lateral response for a step ommand of 10 m

213
 3.2 response for a step ommand of 0.1 rad
3.3 response for a step ommand of 0.1 rad
4.1  and  in ase of an engine failure
4.2 lateral deviation in ase of an engine failure
4.3 heading rate _ in ase of an engine failure

16.7 Results of the Automated Evaluation Pro-

edure
This hapter presents the methodologyindependent evaluation results for the
ontroller design des ribed in the previous hapters. The results relate to the
evaluation s enario and mission dened in Chapter 14. They are obtained by
using the RCAM automati evaluation software. Both overall tra king per-
forman e and inner-loop behaviour of the ontrolled air raft are evaluated
by means of bounds of key variables like tra king error (as performan e mea-
sure), normal a elerations (as omfort riterion), variation in tra k angle (as
robustness measure), and angle of atta k (as safety riterion).

Segment I: the ee t of engine failure

This segment of the traje tory is a straight line of about 16 km length at
onstant altitude. A onstant wind with 10 m/s is assumed, dire ted along the
negative earth-xed axes.

First segment: top view Second segment: lateral deviations

300

100
200

50
lateral deviation [m]

100
xdeviation [m]

0 0
0 a b 1
1 c d 2

100
50

200
100

300
20 18 16 14 12 10 8 6 4 2 0 1 2 3 4 5 6 7 8
yposition (YE) [km] along track distance from point 1 [km]

Figure 16.10: Segment I: the ee t Figure 16.11: Segment II: plane
of engine failure. view of the 3 deg/s turn and lateral
deviations.

A single engine failure is simulated in this segment. Sin e RCAM is a

twin-engine air raft, a single engine failure will result in a signi ant lateral
deviation. Figure 16.10 provides a plane view of the rst traje tory segment.
The given bounds provide an a eptable level of performan e. The stati lateral
deviation should be below an a eptable value of 20 m.

214
 Segment II: the 3 deg/s turn
This segment begins with a 90-degree turn followed by a straight line segment,
all at onstant altitude. The onstant wind is still blowing and hen e be omes
a head wind in the progress of Segment II.

The perfe t following of the required traje tory would require a sudden
hange in the air raft's bank angle, whi h is physi ally impossible and obviously
not desirable for omfort reasons. Hen e deviations from the desired traje tory
are unavoidable. This is shown in Figure 16.11 as a plane view of the lateral
deviations from the required traje tory.

Segment III: the apture of the -6 and -3 degree glideslope

The performan e of the air raft during this segment is measured by the de-
viation from a -6 degree glideslope, whi h hanges to a -3 degree glideslope
afterwards. A maximum deviation of 20 m should not been ex eeded. Further-
more, speed variations should be kept small during the ne essary variations of
the angle of atta k and while the head wind redu es to zero.

In Figure 16.12 the verti al deviations of the air raft following the glideslope
are plotted together with bounds of a eptable behaviour.

Segment IV: the nal approa h with windshear

While on a nal approa h with a glideslope of -3 degree, the ee t of a wind-
shear is onsidered. Figure 16.13 shows the bounds that dene an a eptable
longitudinal response to the windshear together with the verti al deviations
from the desired glideslope.

Third segment: altitude deviations Fourth segment: altitude deviations

30 30

20 20
altitude deviation [m]
altitude deviation [m]

3
10 2 f 3 10
4
g h
0 0

10 e 10

20 20

30 30
16 15 14 13 12 11 11 10 9 8 7 6 5 4 3 2 1
xposition (XE) [km] xposition (XE) [km]

Figure 16.12: Segment III: the ap- Figure 16.13: Segment IV: longi-
ture of the -6 and -3 degree glides- tudinal deviations for the nal ap-
lope and verti al deviations. proa h with windshear and verti al
deviations from the desired glides-
lope.

215
Numeri al Results
Using the riteria given in Chapter 14, omparison indi ators for ea h segment
of the evaluation ight prole have been al ulated using the RCAM auto-
mati evaluation software. These values are given in Table 16.2 together with
an overall average for ea h of the ve riteria. For performan e, perfoman e
deviation,  omfort, and safety, values smaller than one indi ate a eptable
behaviour a ording to the requirements stated in Chapter 14. Power mea-
sures the ontrol a tivities and is not normalized to one. Its value serves for a
relative omparison of ontrollers.

Segm. I Segm. II Segm. III Segm. IV Total

Performan e 0.1270 0.1436 0.2587 0.1781 0.1769
Perf. Dev. 0.0438 0.0184 0.1121 0.0972 0.0679
Comfort 0.5500 1.6295 1.3224 0.5175 1.0049
Safety 0.0051 0.1281 0.0081 0.0736 0.0537
Power 0.0026 0.0104 0.0150 0.0308 0.0147

Table 16.2: Numeri al results of the evaluation pro edure

The table shows that besides omfort during Segment II and III all values
are less than 1 and hen e satisfa tory. It is up to the designer to judge whether
this is an a eptable trade-o or not.
The above diagrams and the numeri al values of the table show that the
designed ontrollers are quite robust with respe t to the investigated parameter
variations. Moreover, the analysis of Se tion 16.6 shows that the ontrollers
remain stable with su iently good performan e even for mass variations up
to 50 000 kg. This shows again that the multi-model approa h is well suited
for designing robust ontrollers.

16.8 Con lusions

Flight Control is not a new and unsolved ontrol appli ation area. Hen e there
is a lot of proven knowledge about using suitable ontroller stru tures and
setting up meaningful spe i ations. But a best possible appli ation-spe i
parameter tuning of ontrol laws is still a time- onsuming trial and error task.
The multi- riteria/multi-model/multi-parameter optimisation approa h as
demonstrated here aims at relieving that tuning task in a transparent and
systemati way. As shown, this approa h results in low-order ontrollers with
good performan e. In parti ular,a quite robust solution resulted, as measured
by the respe tive ben hmark evaluation gure of merit. This indi ates the
usefulness of the multi-model approa h.
Note, that for longitudinal ontrol we intentionally applied an LQR-state
feedba k approa h whi h is realized as omplete output feedba k. This was
done in order to demonstrate the feasibility of using an analyti al synthesis
method within the parameter tuning optimization loop as well. In an industrial

216
appli ation, more likely, another ontroller stru ture would have been hosen.
It is a strength of the MOPS method that it is not bound to a spe i ontroller
stru ture or spe i ontrol law synthesis method.
The multi- riteria/multi-model design approa h is easy to omprehend be-
ause problem formulation and design- y le setup is straight forward in terms
of the design spe i ations. The omputer, not the engineer, takes on the work
of numeri ally tuning the appropriate ontrol design parameters.
The eort related to setting up the design y le, i.e. setting up design
models and ontroller stru tures, is essentially the same as in any other general
ontroller synthesis method. More emphasis than in other methods is put on
setting up the design riteria and this is an engineering appli ation strength of
the MOPS method. For a spe i appli ation area like ight ontrol, generi
design-obje tives as well as problem-spe i obje tives like handling qualites
riteria an be implemented for re-use in an appropriate design environment.
The design y le is exe uted intera tively. Synthesis of new ontroller pa-
rameters is done automati ally by the omputer, whereas the de ision making
for appropriate models, riteria, or demand values, requires an intera tion by
the design engineer.
Computation time for optimization may be in the order of minutes, depend-
ing on the problem to be solved.
The a tual eort ne essary for performing a redesign after a major design
de ision hange, depends on what has been hanged: models, ontroller stru -
tures, or performan e spe i ations. The essential feature is that the multi-
riteria/multi-model design approa h is most exible and systemati to deal
with su h hanges.

217
17. An Eigenstru ture Assignment
Approa h (1)

Lester Faleiro1 and Roger Pratt 1

Abstra t. This hapter is a des ription of the use of one parti ular
methodology of eigenstru ture assignment on the Robust Control
Air raft Model (RCAM) hallenge as set out in [145. Emphasis
is put on the general methodology in order to on entrate on the
similarities between traditional ontrol system design methods and
eigenstru ture assignment. The results show that although the nal
ontroller in this ase is not the ideal one, the design pro ess is
dened learly enough for eigenstru ture assignment to be used as
a design tool in a wider design pro ess.

17.1 Introdu tion

Various methodologies of eigenstru ture assignment are detailed in hapter 3,
and some of them are ustilised in this hapter to produ e one possible solution
to the design problem. The design pro ess that is presented here is not the
ideal one by any means, but the intention is to demonstrate the attributes that
eigenstru ture assignment has as a design tool, and the potential results it an
deliver.

The se tions of the hapter are set out in the order in whi h the design
pro ess an be arried out. However, it must be remembered that the pro ess
is an interative one, as will be ome lear.

17.2 The Sele tion of the Controller Ar hite -

ture for the RCAM Problem
For most xed-wing air raft, there is very little ross- oupling between the
longitudinal and lateral dynami s. Consequently, it was de ided that two on-
trollers would be used; one for the longitudinal dynami s, and one for the
lateral dynami s.

218
 SPPO Phugoid Displa ement
Eigenvalue 0:8303  1:1069i 0:0114  0:1264i 0
 0.6 0.0898 -
!n 1.3837 0.1269 -
q 0.0136 0.0002 0
 0.0098 0.0016 0
u 0.0144 0.1216 0
w 0.9430 0.0175 0
z 0.3320 0.9924 1

Table 17.1: Eigenve tors of the longitudinal open-loop system

17.2.1 Longitudinal ontroller

The open-loop longitudinal dynami s for RCAM at a nominal trim ondition
(mass at 120 tonnes, horizontal entre of gravity at 23% of mean aerodynami
hord (MAC) and verti al entre of gravity at its lowest point) an be deter-
mined using GARTEUR software. The eigenstru ture of this system is shown in
Table 17.1. For the sake of simpli ity, a tuator dynami s will be omitted. The
dynami s ontain the traditional Short Period Pit hing Os illation (SPPO) and
phugoid modes, and an additional verti al displa ement mode, oupled only to
the verti al displa ement state.

VA z T TH
VA z
-
command +
error
Llon +
+
Aircraft and
actuation

Trim - Klon
conditions + q nZ VA wV z

VA z

Figure 17.1: Longitudinal losed-loop system stru ture

The hosen design task was to regulate pit h rate, verti al a eleration,
airspeed, verti al velo ity and verti al displa ement( [q nz VA wv z ). Based on
the design requirements to tra k hanges in verti al displa ement and airspeed,
these two variables were hosen as tra ked outputs. The full linear losed-loop
system an now be depi ted as shown in Figure 17.1. There are two main
omponents to the stru ture:
1. The ve feedba k signals are used to regulate the air raft. This is done
by multiplying the error between the output signals and the trim onditions
1
Department of Aeronauti al and Automotive Engineering and Transport Studies, Lough-
borough University, Loughborough, Lei estershire LE11 3TU, United Kingdom.

219
by the stati gains in the matrix Klon, whi h produ es taileron and throttle
signals to return the air raft to the trim ondition.

2. The errors between the referen e signals and their respe tive outputs
are integrated and fed through a gain matrix, Llon , whi h will ensure that the
error between the referen e signal and the output signal is always zero.

17.2.2 Lateral ontroller stru ture

As with the longitudinal air raft model, the lateral linear dynami s for the given
trim ondition an be determined using GARTEUR software. The eigenstru -
ture of this system is shown in Table 17.2. Again, the modes of the open-loop
system are onventional.

Roll Spiral Dut h Roll Heading ylat

Eigenvalue -1.3017 -0.1837 0:2360  0:5954i 0.00 0
 - - 0.3684 -
!n - - 0.6405 -
p 0.2177 0.0008 0.0170 0.0008 0
r 0.0145 0.0005 0.0049 0.0005 0
 0.1669 0.0043 0.0264 0.0043 0
0.0111 0.0028 0.0076 0.0028 0
vB 0.0138 0.0472 0.8901 0.0472 0
ylat 0.9614 0.9989 0.4546 0.9989 1

Table 17.2: Eigenve tors of the lateral open-loop system

There are eight measured outputs available in the lateral dynami s of RCAM.
Only ve of these are ne essary to implement su ient ontrol over the four
modes des ribed in the open-loop dynami s. It was hosen to regulate the
hanges in sideslip angle, roll rate, yaw rate, roll angle, tra k angle and lateral
deviation ( [ p r   ylat). Based on the design requirements to tra k hanges
in heading rate and lateral displa ement, the tra ked outputs were hosen as
roll angle (whi h is dire tly related to heading rate) and lateral displa ement.
All other lateral demands an be translated into a ombination of these.

The linear losed-loop system an be depi ted s hemati ally as shown in

Figure 17.2. Note that ylat is used both as a regulation feedba k signal and as
an output signal. For normal operation, the error from  is given by dieren e
between the initial  and  output. However, when a turn rate is demanded,
the desired steady-state  has to be altered to prevent the regulator from
attempting to keep it at its initial ondition. This is done by integrating a
demand to get a desired at any time step and simply adding that value to the
initial . Some aspe ts of the stru ture des ribed here were added on during
the design pro edure itself, and the pro ess is inevitably iterative.

220

c Va Va
c = atan -------------
c
g

ylat c

- ylat A R

error + Aircraft and
+
Llat actuation
+
Trim - +
conditions
pr
error p r
+ Klat
trim
error
error
-
ylat

- +
c + 0.2 + additional

Figure 17.2: Lateral losed-loop system stru ture

17.3 The Translation of RCAM Design Criteria

into Method Dependent Obje tives
The RCAM hallenge design spe i ations are detailed in [145. The simple
eigenstru ture assignment design synthesis performed for the RCAM design
only requires two sets of inputs. The rst are the system matri es A, B and
C. The se ond onsists of a set of desired losed-loop eigenvalues and a set of
desired losed-loop eigenve tors ( d and Vd ). This se tion des ribes how the
various design spe i ations an be transformed into desired eigenvalues and
eigenve tors.

17.3.1 Performan e riteria

Performan e requirements are the most signi ant requirements in RCAM. The
riteria are omposed of rise time and settling time spe i ations, overshoot
limits and ross- oupling limitations. The following se tions relate to the re-
quirements des ribed in [145, and only spe i ations pertinent to the use of
eigenstru ture assignment have been in luded.
Both the longitudinal and the lateral systems ontain se ond order modes,
and all the tra king requirements are given for step input onditions. The
response of a se ond order system to a step input is well do umented. The
dierent spe i ations for rise time and overshoot for a parti ular system an

221
thus be transformed into eigenvalues, and any states that should not exhibit a
mode an then be de oupled from that mode in the eigenve tors (see hapter
3 for a fuller explanation).

Lateral deviation
The tra king requirement for lateral deviation provides a minimum limit for
the natural frequen y and damping ratio of the se ond order modes onne ted
with ylat. This allows a ertain exibility in the eigenstru ture assignment
pro edure, as an improved de oupling solution may be obtained by allowing
the eigenvalues to roam within this limit.

Altitude response
This is a similar spe i ation to the lateral deviation requirement. Again, the
overshoot limit means that verti al displa ement should not be oupled into
modes that have a damping ratio of less than 0.7. The verti al deviation lim-
its for low level ight is more di ult to in orporate into the eigenstru ture
assignment design. Essentially, this an be ta kled by simply improving ro-
bustness as mu h as possible, so that the hange in altitude will not result in a
large deviation of the performan e of the air raft from the nominal losed-loop
system.

Roll angle response

This spe i aiton relates to engine failure. The only thing that an be done in
eigenstru ture assignment is to have a relatively qui k response to ylat devia-
tions, so that in a regulation situation, roll angle will not have the opportunity
to be ome large enough to violate the given limits. Additionally, de oupling
roll angle from modes that might be ome ex ited during an engine failure,
su h as the Dut h Roll, should prevent large roll angles from o urring during
regulation.
The requirement to de ouple sideslip from roll motion an be in orporated
into the eigenstru ture assignment pro edure easily. Lateral velo ity hanges
(and hen e hanges in sideslip for a onstant forward velo ity) are de oupled
from the modes relating to rolling motion by inserting a zero into the relevant
eigenve tor element. This will produ e a o-ordinated turn.

Airspeed response
The requirements for airspeed tra king an again be transformed into bounded
eigenvalues for the modes oupled to airspeed. The limit on airspeed deviation
in the presen e of a step disturban e in wind velo ity an only be addressed
by having su ient damping on the mode involving velo ity to insure that it
is redu ed to allowable levels in the time required. The requirement to have no
steady state velo ity error an be easily dealt with by using the velo ity error
integrator des ribed in the longitudinal ontroller stru ture.

222
Cross- oupling between airspeed and altitude
This requirement is a de oupling spe i ation, and an be in orporated into
the desired eigenve tors of the longitudinal ontroller. As shown in Table 17.3,
u has been de oupled from z tra k and w has been de oupled from VA tra k.

17.3.2 Robustness riteria

Unfortunately, the method do umented here annot yet be used to design for
exa t parameter variations. Instead, the obje t is to make the system as insen-
sitive to parameter variation as possible, thus attempting to preserve nominal
stability and performan e whilst arbitrarily improving robustness.

17.3.3 Spe i ation of eigenstru ture

A set of initial estimates, based on the desired performan e, robustness and
stability of the losed-loop air raft an now be proposed. Note that this set is
not unique, and it is very likely that the desired eigenstru ture will have to be
updated before a suitable design an be found.

Longitudinal mode spe i ation

In keeping with onventional air raft behaviour, the losed-loop longitudinal
system an be assigned three modes: the SPPO, the phugoid and a verti al
displa ement mode. In addition to these, the eigenstru ture assignment pro-
edure augments the open-loop system matri es with the two integrated error
states of the outer-loop tra king. The eigenstru ture that is hosen for the rst
design is shown in Table 17.3.

SPPO Phugoid z VA tra k z tra k

Eigenvalue 0:8  0:8i 0:15  0:15i -0.3 -0.4 -0.5
q x x x x x
 x x x x x
u 0 x 0 x 0
w x 0 x 0 x

R
z x x x x x
V
R A
x x x x x
z x x x x x

Table 17.3: Desired eigenstru ture of the longitudinal losed-loop system

Lateral mode spe i ation

As with the longitudinal system, there are traditional modes of behaviour that
an be dened for the losed-loop system. These in lude a Roll mode, a Spiral
mode and a Dut h roll. Additional modes that we an now dene are a heading

223
mode and the two integral tra king modes. An initial desired eigenstru ture is
shown in Table 17.4.

Roll Spiral Dut h Roll Heading ylat  ylat

tra k tra k
Eigenvalue -4.4 -0.2 0:182  0:157i -0.13 -0.55 -1.5 -0.5
p x x 0 x x x x
r x x x x x x x
 x x 0 x x x x
x x x x x x 0
vB 0 0 x 0 0 0 x
ylat
R x x x x x x x

R  x x 0 x x x x
ylat x x 0 x x x x

Table 17.4: Desired eigenstru ture of the lateral losed-loop system

17.4 The Des ription of the Design Cy le

Only a few tools are required in order to produ e and analyse an eigenstru ture
assignment design that works. However, the pro ess is still a di ult one.
This is not to do with the tools themselves. Rather, it has to do with the
interpretation of the results of the analysis and the onsequent de isions that
the designer must make in spe ifying an updated eigenstru ture.

17.4.1 Initial synthesis

The rst design synthesis is a simple one. All that is required is the robust
eigenstru ture assignment algorithm. There are two basi sets of inputs to this
program. The rst is a set of system matri es, A, B , C and Ca. The last of
these is a matrix that ontains the rows ofC whi h des ribe the outputs that are
to be tra ked. The se ond is the desired eigenstru ture d and Vd , The values
of the initial estimates for these are given as method dependent obje tives in
Table 17.3 and Table 17.4. These are simply entered into a program to produ e
the stati gain matri es K and L.

17.4.2 Intermediate analysis

The intermediate analysis of the ontroller takes up most of the time in this
design pro ess. This is done using a ombination of linear simulation and
eigenstru ture analysis to ensure that performan e goals are satised, and some
sensitivity analysis to determine how robustness goals are being met.

The longitudinal ontroller

Eigenstru ture analysis an be used to examine the system. The losed-loop
eigenstru ture of the initial design is given in Table 17.5. The shaded ele-

224
ments show that the de oupling that was spe ied in the desired losed-loop
eigenve tors has been attained exa tly.

 1:4559 0:1129 0:0025 0:0241 0:0075

  0:0008 0:0009

Klon = 0:3091 0:0300 0:0123 0:0006 0:0020
Llon = 0:0017 0:0002

SPPO Phugoid z (Taileron) (Throttle) VA tra k z tra k

 0 :8  0 :8 i 0:150:15i -0.3 -5.0285 -0.8875 -0.4 -0.5
 0.71 0.71 - - - - -
!n 1.13 0.21 - - - - -
q 0.0086 0.0001 -0.0002 0.0654 -0.0022 -0.0006 -0.0003
 -0.0076 0.0005 0.0005 -0.0130 0.0025 0.0015 0.0005
u 0 -0.0457 0 0.0155 -0.0051 -0.2483 0
w 0.8279 0 -0.0428 -0.9910 0.6469 0 -0.1782
z -0.4198 0.2024 0.2871 -0.0095 -0.5061 0.2763 0.4400
t -0.0026 0.0002 0.0004 0.1150 -0.0045 0.0015 0.0014
th -0.0012 0.0005 0.0004 -0.0004 -0.0012 0.0053 0.0009
(V ) 30400 Gain margin -4.3 dB, 9 dB Phase margin  38 degrees

Table 17.5: Eigenstru ture of the longitudinal losed-loop system

The intera tion of the tra king ommands with the outputs an be followed
qualitatively by using the losed-loop input oupling ve tors. The oupling
between a ommanded input and the modes of the system is given by:

Wlon Blon Llon (17.1)

These ve tors are shown in Table 17.6. These oupling ve tors show us that
when there is a demand in VA , the Phugoid, SPPO, z, z tra k and VA tra k
modes will be strongly ex ited. The other modes will be involved, but not to
as great an extent. Looking ba k at Table 17.5, when these modes are ex ited,
it will ause an ex ursion in both forward velo ity and in verti al displa ement.

VA demand z demand
SPPO 0.3901 0.4595
Phugoid 0.9686 0.0430
z 4.0196 1.4312
(Taileron) 0.0680 0.0792
(Throttle) 0.7187 2.0467
VA tra k 0.3574 0.0486
z tra k 1.7478 3.1458

Table 17.6: Input oupling of the longitudinal losed-loop system

When a hange in z is demanded, the SPPO, the verti al mode and the z
tra k mode are the most involved. These modes do alter the z state, but are not
very dominant in the forward velo ity state. Thus, our ursory examination of
VA is de oupled
the eigenstru ture of the system indi ates that it is likely that
from a z demand, but z does not appear to be de oupled from a VA demand.
This evaluation of the eigenstru ture an be tested by using linear simu-
lation of step demands on the tra ked variables. Figure 17.3 shows the time
response of this system to a step demand in rease of 13 m/s in VA . Note
that, as predi ted by eigenstru ture analysis, there is a large ex ursion in z. A

225
step ommand of 30m in rease in z produ es the results shown in Figure 17.3.
Again, as predi ted, VA has been de oupled well enough from a demand in z
to satisfy the design spe i ation.

15 35

VA 30
10
25
5
z

m, m/s
wv
m, m/s

20

0 15

10
5

z 5 wv
10
0
VA
15 5
0 10 20 30 40 50 0 10 20 30 40 50

Time (seconds) Time (seconds)

Response to 13 m/s step in VA Response to 30m step in z

Figure 17.3: Step responses of initial longitudinal system

It is now up to the designer to alter the design parameters on the basis

of this analysis. Although this initial system has satised most of the design
spe i ations, the verti al a eleration involved is outside the allowable limits,
espe ially with a z demand. A glan e at the time responses shows that this is
likely to be the result of an under-damped SPPO mode. The damping on the
desired SPPO mode an now be altered by the designer to obtain the desired
ee t.

The lateral ontroller

The lateral system an be analysed in exa tly the same way as the longitudinal,
as ontroller stru tures are similar. Again, it is important to take designer
experien e into a ount during the design pro ess.
An initial ontroller for this system was designed and analysed, and was
found to be useful for linear simulations whi h satisfy the step demand require-
ments detailed in [145.

17.4.3 Robustness and goal attainment

One of the most evident problems with eigenstru ture assignment is that any
small alteration of the eigenvalues alters the ve tor spa e from whi h the a hiev-
able eigenve tors an be sele ted. This means that although it is easy for the
designer to hoose arbitrary eigenvalues to produ e desired responses, it is not
possible to examine the ee t of this hoi e on the nal eigenve tors of the
losed loop system. Ea h hoi e of eigenvalues will ae t the many parameters
by whi h the nal design an be measured. One su h parameter is stability
robustness. A set of design goals an be used to spe ify desired stability ro-
bustness, whi h an be des ribed by the minimum singular value of the return
dieren e matrix (RDM) at the input and output of the system. These values
an be translated into onservative gain and phase margins. The exa t rela-
tions are do umented in [152. Goal attainment oers a method of using any
freedom inherent in the system to sear h the allowable set of solutions itera-
tively for a design that satises as many of these robustness goals as possible.

226
A simple diagrammati al approa h to the goal attainment pro edure, as well
as a fuller explanation of the mathemati al ba kground involved an be found
in [95.
To preserve the desirable qualities of a system designed by eigenstru ture
assignment, goal attainment was implemented with an internal eigenstru ture
assignment algorithm. The pro edure begins with an arbitrary set of eigen-
values, and produ es a feedba k gain matrix as a solution to a robust eigen-
stru ture assignment problem spe ied by the designer. This feedba k gain
matrix, K, is then used to evaluate the nearness of the stability robustness
measures to the desired robustness goals. The desired eigenvalues and/or de-
sired eigenve tors are then altered a ordingly, and the eigenstru ture assign-
ment pro edure is repeated until a solution has been obtained. This pro edure
is depi ted s hemati ally in Figure 17.4 and the resulting ontrollers are shown
in Table 17.7.

Algorithm to Have
compute NO goals been YES K and L (final)
updated parameters attained?
A,B,C
V Initial V
d d d d
Goals for performance
Eigenstructure and stability robustness
Assignment
Procedure
Designer interaction
K and L

Obtain performance
and stability
measures Goal attainment
algorithm

Figure 17.4: S hemati of goal attainment pro edure

These ontrollers have been designed using only eigenvalue perturbations.

Comparing them with that given in Table 17.5, it an be seen from both the
ondition numbers of the eigenve tor matri es and the multivariable stability
margins that goal attainment an be used with eigenstru ture assignment to
produ e a system with improved robustness.

17.5 Analysis of the Resulting Controller

The ontroller was analysed with respe t to the a hievement of the design
spe i ations. Where step responses have been demanded, these are performed
on the non-linear simulation. Unless otherwise spe ied, the results are for the

227
  
3:088 0:269 0:008 0:048 0:014
Klon =
1:441 0:131 0:012 0:018 0:07
 
2:435 5:165 4:599 19:571 8:077 0:050
Klat =
0:005 2:232 0:001 3:755 0:437 0:005
   
0:0008 0:0019 0:3006 0:0028
Llon = Llat =
0:0013 0:0008 0:1888 0:0003
Longitudinal Lateral
First design (V ) 30400 34730
(V ) 8439 30640
Gain margin -4.8 dB, 11.3 dB -3.7 dB, 6.7 dB
Phase margin 43 degrees 31 degrees
No. of Iterations 200 100

Table 17.7: Chara teristi s of the nal ontroller designed with goal attainment

losed-loop air raft at nominal ondition (120 tonnes, horizontal 23% mean
aerodynami hord and verti al 0% mean aerodynami hord entre of gravity;
time delay at 0.05s). Any design spe i ations that are satised have not been
do umented, but all the required tests are shown in graphi al form in Figure
17.6.

17.5.1 Performan e riteria

1
Deviation (metres)

0.5
Lateral demand
Lateral disturbance
0

0.5
0 20 40 60 80 100

Time (seconds)

Figure 17.5: Lateral deviation performan e

Lateral deviation
The lateral deviation of the air raft from an initial ondition, 1m o the desired
traje tory, is shown in Figure 17.5. A redu tion of lateral deviation to 10% of
its original value o urs in less than 8 se onds, but also results in an overshoot
of 25%. This is not a eptable performan e. Earlier designs did not have this

228
bad performan e, but it was required in order to ensure that the air raft did
not stray outside the allowable bounds in the event of an engine failure.

Flight path angle response

This response has an overshoot of only 1.6%, but does not respond as qui kly
as desired to ommands in ight path angle. This situation annot be easily
remedied using the eigenstru ture assignment ontroller, as it has been designed
to solve independent z demand and VA demand situations. A demand implies
a on urrent VA and wV demand.

Roll angle response

The design spe i ations require that a small roll angle be maintained after
engine failure to keep sideslip minimised. However, the eigenstru ture assign-
ment design is su h that roll angle is minimised instead. As an be seen in
Figure 17.5, this results in a steady sideslip of about 3 degrees. It also invali-
dates the remaining design riteria of not overshooting 50% of the steady roll
angle on engine start-up.

Cross oupling between airspeed and altitude

Although ross oupling was insigni ant during the early designs, later alter-
ations have aused a 1.4 m/s deviation (desired at less than 0.5 m/s) in VA for
a demand of 30 metres in height ( ). z

17.5.2 Robustness riteria

It was found that for a time delay of 50 ms, the system had robustness over all
other parameter variations spe ied ( hanges in entre of gravity and mass).
With the maximum time delay of 100 ms, in a small part of the parameter
envelope, namely at aft and high entre of gravity with a mass of greater
than 145 tonnes, the system qui kly began to lose robustness, and eventually
be ame unstable. However, it must be mentioned that this o urred during the
3 degree/s turn, whi h has already been shown to be a bad segment for this
parti ular ontroller.

17.5.3 Ride quality riteria

These riteria relate to passenger omfort under normal manoeuvres. As su h,
the results from the evaluation simulation have been taken as representative.

Maximum verti al a eleration

This is desired to be less than 0:05g. Figure 17.7 shows that this value is
violated twi e. The less harsh violation is during the wind shear, where we
would expe t an un omfortable ride. However, the worst violation, at 0:4g,

229
 30 10

PSI (degrees)
9
25
Z (metres)

8
7
20
6

15 5
4
10 3

2
5
1

0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30

Response to 30m step in Z Response to 10 deg step in heading angle

PHI / BETA (degrees)

3 10
GAMMA (degrees)

8
2.5
6

2
4

2

1.5 0

2
1 4

0.5
6
8

10
0
0 5 10 15 20 25 30 35 40 0 50 100 150 200

Flight path angle step response Roll and sideslip during engine failure
and re-start
3 2
PHI (degrees)

PSIDOT (degrees)

2 1.5

1 1 failure
0 0.5

1 0

2 0.5

3 1
restart
4 1.5
0 5 10 15 20 25 30 35 40 0 50 100 150

Roll angle under moderate Heading rate during an engine failure

turbulence
6
1.2
4
1 2
VA (m/s)
VA (m/s)

0
0.8
2

0.6 4

6
0.4
8

0.2 10

12
0
0 10 20 30 40 50 60 70 80 14
0 10 20 30 40 50 60 70 80

Step response of airspeed Airspeed deviation response to a

wind step in the direction of UB
14
VA, Z (m/s, m)

30
VA, Z (m/s, m)

12
25
10
VA 20
8
z
6
15
4
z 10
2 VA
0 5

2 0
0 10 20 30 40 50 60 70 80
0 10 20 30 40 50 60

Response to a 13 m/s VA demand Response to a 30m Z demand

Figure 17.6: Time responses of nal system

230
 omes during the steady turn. Of ourse, this is to be expe ted, and has already
been explained.

0.1 0.2
0.05
0.15
0

0.05 0.1

ny(g)
nz (g)

0.1
0.05
0.15

0.2 0

0.25 0.05
0.3
0.1
0.35

0.4
0 100 200 300 400 500 0 100 200 300 400 500

Time (seconds) Time (seconds)

Vertical acceleration Lateral accleration

Figure 17.7: A elerations during evaluation

Maximum lateral a eleration

This should be less than 0:02g. Figure 17.7 shows that this requirement is
violated twi e. The rst time is during the engine failure, whi h is onsidered
allowable for this emergen y situation. The se ond time is during the steady
turn, where large a elerations of 0:08g o ur.

17.5.4 Safety riteria

Only two of the safety riteria need to be examined with respe t to the way in
whi h the design has ae ted them.
The roll angle over the evaluation simulation for the nominal system with
a 21% mean aerodynami hord verti al entre of gravity are shown in the rst
plot in Figure 17.8. The required roll angle limit is 30 degrees. This is ex eeded
by the air raft with a high entre of gravity. However, it an be seen that this
is always aused by the bad performan e of the system during the steady turn.
Without a lateral deviation ommand, the response of the system would be as
shown in the se ond plot in Figure 17.8, but the lateral deviation regulation of
the system is introdu ing another 7 to 15 degrees of roll angle.

25
35

30 20
(degrees)
(degrees)

25

20 15

15
10
10
5
5
0

5 0
10

15 5
0 50 100 150 200 250 300 350 0 10 20 30 40 50 60

Time (seconds) Time (seconds)

Roll angle during evaluation Roll angle during independent turn

Figure 17.8: Roll angle hara teristi s

231
 RMS rate (deg/s) Maximum in
design spe s (deg/s)
Aileron 15.5 8.3
Tailplane 4.3 5
Rudder 0.6 8.3
Throttle 0.32 0.24

Table 17.8: Satisfa tion of the maximum rate requirements

17.5.5 Control a tivity riteria

The requirements are that mean a tuator rates (taken as RMS a tuator rates
here) should be less than 33% of the maximum rates. Figure 17.9 shows the
a tuator movements under `moderate' turbulen e onditions. Note that the
aileron exhibits ex essive movement. This is again due to the fa t that it is
trying to regulate lateral deviation too qui kly. This an be seen more vividly
in Figure 17.9. The dependent axes on these plots are s aled to the minimum
and maximum rates of the a tuator on erned. The aileron onstantly rate
saturates. The remainder of the a tuators perform well.

17.6 Results of the Automated Evaluation Pro-

edure
This se tion is based on the evaluation mission and s enario dened in the
RCAM manual [145. Figure 17.10 shows the time plots of the air raft with
the eigenstru ture assignment ontroller. The dotted lines indi ate performan e
bounds that should not be ex eeded. Four runs are displayed on ea h plot to
determine a measure of the robustness of the system under parameter variation.
The only segment that indi ated possible problems with the ontroller was
the se ond segment, the 3 deg/s turn. These problems and the reasons for their
o urren e have already been analysed during linear simulation.

Segm. I Segm. II Segm. III Segm. IV Total

Performan e 0.0707 0.1454 0.3427 0.1707 0.1824
Perf. Dev. 0.0324 0.0121 0.0828 0.2120 0.0848
Comfort 0.8864 3.3991 1.2201 0.4523 1.4895
Safety 0.0159 0.0749 0.0122 0.1013 0.0511
Power 0.0079 0.0098 0.0148 0.0329 0.0163

Table 17.9: Numeri al results of the evaluation pro edure

Quantitative measures of the ontroller are shown in Table 17.9. All the
values in the table have been normalised to maximal allowable bounds ex ept
for power. Thus, ex eeding a value of unity indi ates that the relevant bounds

232
 Actuation (degrees)
A R
20

15 TH
10

10

15 T
20
100 110 120 130 140 150

Actuator movements during evaluation under moderate turbulence

25 25

20 20
A rate (deg/s)

R rate (deg/s)
15 15

10 10

5 5

0 0

5 5

10 10

15 15

20 20

25 25
100 110 120 130 140 150 100 110 120 130 140 150

15 1.5
TH rate (deg/s)
T rate (deg/s)

10 1

5 0.5

0 0

5
0.5

1
10

1.5
15
100 110 120 130 140 150 100 110 120 130 140 150

Actuator rates during moderate turbulence

Unless specified, the independent axis describes time in seconds

Figure 17.9: A tuator hara teristi s

First segment: top view Second segment: lateral deviations

300

100
200

50
lateral deviation [m]

100
xdeviation [m]

0 0
0 a b 1 1 c d 2

100
50

200
100
300
20 18 16 14 12 10 8 6 4 2 0 1 2 3 4 5 6 7 8
yposition (YE) [km] along track distance from point 1 [km]

Segment I: the effect of engine failure Segment II: lateral deviations during the
3 deg/s turn
Third segment: altitude deviations Fourth segment: altitude deviations
30 30

20 20
altitude deviation [m]

3
altitude deviation [m]

10 10
2 f 3
4
g h
0 0

10 e 10

20 20

30 30
16 15 14 13 12 11 11 10 9 8 7 6 5 4 3 2 1
xposition (XE) [km] xposition (XE) [km]

Segment III: vertical deviations from the Segment IV: vertical deviations from the
desired glideslope desired glideslope

Figure 17.10: Evaluation pro edure results

233
has been violated.
For the eigenstru ture assignment ontroller, almost all the values in this ta-
ble are adequate. The large values for omfort are due to the importan e pla ed
on ensuring good performan e through a urate tra king of the desired traje -
tory. Thus, in its `eagerness' to keep to the desired traje tory, the air raft uses
short, sharp movements that ause a elerations to ex eed the re ommended
design bounds.

17.7 Con lusions

17.7.1 Requirements for the designer
This hapter has attempted to provide a omprehensive appli ation of sim-
ple eigenstru ture assignment to the RCAM design hallenge problem. The
progression of a ontroller design for RCAM involves the need for two things:
1. The designer needs to have a good knowledge of air raft's dynami s.
2. The designer needs to understand the way in whi h the air raft's eigen-
stru ture relates to these dynami s. This knowledge is required so that, as with
lassi al design, elements in the desired eigenstru ture an be altered with a
knowledge of the onsequen es of that alteration. It has been demonstrated in
this do ument that it is fairly easy to understand, at least qualitatively, the
onne tion between eigenstru ture and system behaviour.

17.7.2 The design pro ess

On e the knowledge des ribed above had been gained for RCAM, design spe -
i ations were used to produ e a simple ontroller stru ture. A desired eigen-
stru ture was then produ ed and used to nd an initial ontroller.
This ontroller was then analysed and altered until a nal ontroller was
produ ed. Along the way, goal attainment was used to improve ea h su essive
design. The entire design pro ess is shown in Figure 17.11. The on ept of
eigenstru ture was used throughout the pro ess in some form or other (analysis
and assignment). Nevertheless, as the diagram shows, the designer intera ts
with the pro ess throughout.
The solid boxes represent pro esses. These are either omputer programs,
or sub-pro esses in the design that are performed using eigenstru tural tools,
as des ribed in the relevant se tions. The dotted boxes represent the initial de-
sign spe i ation as produ ed by the designer, based on her/his knowledge and
experien e. The variables in these boxes an be altered as the design pro ess
progresses. For RCAM, it was only ne essary to alter the desired eigenstru -
ture, and no hanges to the ontroller stru ture were required during the design.
The worst aspe t of the nal ontroller was found to be the behaviour of the
air raft when subje ted to both a roll demand and a lateral deviation at the
same time. The aileron response to these on urrent ommands is una ept-
able, and should be altered. The reason for this response was a re-design to
a ount for large variations in lateral deviation during an engine failure. It is

234
 RCAM information Design Process
Open-loop system Controller structure
analysis
Design specification Desired V
d d
Eigenstructure
A, B, C, Ca Assignment
Procedure
K and L

Intermediate
analysis

Design
specification NO
met?

YES
Goal attainment
Goals for performance Detailed final
and stability robustness analysis
K and L (final)

Non-linear
simulation

Figure 17.11: S hemati of the design pro ess

235
thus likely that the best way to ounter the problem is to design a separate
ontroller for an engine failure ase.

17.7.3 Advantages of eigenstru ture assignment

When ompared with lassi al methods, one advantage of using eigenstru ture
assignment is in the time taken to produ e a robust design. Classi al methods
an produ e a system with good performan e, but robustness is a matter of te-
dious iteration. With eigenstru ture assignment, the designer is able to spe ify
performan e and some robustness limitations during the design pro ess.
As with other modern methods, another advantage is in the use of all the
freedom that is not otherwise utilised in lassi al design to produ e a more
robust system. The fort e of eigenstru ture assignment, though, is in being able
to use some of this freedom to de ouple modes from states. Thus, the method
is primarily a performan e improvement method, and within that, robustness
an also be improved.
The ontroller stru ture is a simple gain matrix. Again, this is omparable
to lassi al methods, where the relation between the outputs of the system and
the a tuation demanded is visible to the designer, who is used to examining
stati gains as part of a design pro edure. This also eases maintenan e and
implementation of the ontroller.

17.7.4 Disadvantages of eigenstru ture assignment

Although the method is able to take a ount of any available freedom in the
losed-loop system, it is unable to a ount for the unassignable modes. These
in lude the (n max(m; p)) unassignable modes and the limitation of tra king
min (m; p) outputs (see hapter 3). This means that it might be ne essary to
insert the algorithm into a goal attainment routine to simply assure that the
system will be stable. These optimisation routines may take a lot of time to
onverge. Additionally, the result of the optimisation depends on the initial
ondition. Thus, altering the initial ondition will result in a dierent on-
troller.
Un ertainties in this pro ess mean that rather than using eigenstru ture
assignment on its own as a omplete design pro ess, the method is best used
as part of an array of tools in analysis and synthesis of a system.

17.7.5 Improvements to eigenstru ture assignment

The method des ribed here has been extended to the addition of dynami
ompensation. This was unne essary for the RCAM ase, and should only be
used if the system annot be designed with a simple stati gain.
As eigenstru ture assignment is a very exible methodology, it should be
possible to extend it to improve the ability to tra k more outputs and to de-
ouple inputs from outputs, rather than modes from outputs. The idea behind
these improvements is that the method should eventually be extremely visible

236
and exible enough for lassi al ontrol designers to be able to understand how
it relates to lassi al ideas.

17.7.6 Summary
Overall, the pro ess was su essful for use with the RCAM design hallenge. It
was possible to use the majority of the design spe i ations given in the RCAM
manual to produ e an a eptable ontroller in linear simulation. It was shown
that with this pro ess, it was possible to onstru t and implement a simple
system that had good performan e and robustness hara teristi s. In addition,
the designer was able to ontrol the trade-o between the two by using their
intuition to alter the desired eigenstru ture.
The only disappointment in this ase was that the same ontroller was not
able to guarantee good performan e for both roll ommand following and lateral
displa ement orre tion, although this was an anomaly for this parti ular ase
where the two requirements were in strong oni t.
It has been shown that eigenstru ture assignment has a simple ontroller ar-
hite ture, a design y le that involves the designer to a large extent, a solution
that satises most of the design spe i ations, and a simple implementation due
to the visibility of the method by lassi al designers. Although the this parti -
ular utilisation of eigenstru ture assignment as a full design method does not
provide all the ne essary answers that designers using lassi al ontrol are look-
ing for, the pro edure is ertainly an useful tool in the analysis and synthesis
of air raft ontrol design systems.

237
18. An Eigenstru ture Assignment
Approa h (2)

Jess M. de la Cruz , Pablo 2Ruiprez

1 2

and Joaqun Aranda

Abstra t. In this hapter the Eigenstru ture Assignment Method
is applied to the Resear h Civil Air raft Model (RCAM) ben hmark
problem. The design is done by making use of the lassi al approa h
whi h onsists of splitting up the ontroller into two parts, a longitu-
dinal and a lateral ontroller, and in using the standard inner-outer
loop ontrol stru ture. This method allows us to dire tly satisfy
spe i ations in terms of transient response and modes de oupling.
However, it does not ope dire tly with system un ertainties. In
order to ope with robustness, measurements of multiloop gain and
phase margins are used in the hoi e of a robust eigenstru ture.
Both hannels make use of a onstant gain matrix in the inner
loop and a s alar gain in the outer loop. In spite of the ontroller
simpli ity, good performan e and robustness results are obtained,
although robustness may be in reased by means of gain s heduling
with respe t to airspeed.

18.1 Introdu tion

The eigenstru ture te hnique has been widely applied to the design of ight
ontrol systems and it is a well-known fa t that its use requires an in-depth
knowledge of the system to be ontrolled [11. Although this method is not
intended to deal with robustness, many robusti ation pro edures have been
proposed [155. Here the eigenstru ture is hosen so that good multiloop gain
and phase stability margins are obtained.
The hapter is organised in the following way. Se tion 18.2 des ribes the
ontroller stru ture used. The ontroller has been de oupled into the longi-
tudinal and the lateral hannels. Both have been designed a ording to an
inner/outer loop ontrol stru ture. The inner loop ontrollers are designed
following the eigenstru ture method. A onstant gain matrix is used in both
hannels. The outer loop only uses a s alar gain that is al ulated by means
of the root lo us method.

1
Dpt. Informti a y Automti a. Fa ultad de Cien ias Fsi as. Universidad Complutense.
28040-Madrid. Spain. (Funded by proje t CICYT TAP94-0832-C02-01).
2
Dpt. Informti a y Automti a. Fa ultad de Cien ias. U.N.E.D. 28040-Madrid. Spain.

238
 Se tion 18.3 deals with the way the RCAM design riteria are translated
into the desired eigenstru ture.
Se tion 18.4 des ribes the design y le. The linear model of the plant is
analysed and the most appropriate eigenstru ture is hosen. Next, the feed-
ba k ontroller is obtained and the performan e of the losed-loop system for
the linear models is analysed. Multiloop gain and phase margins are used to
measure robustness and to guide eigenstru ture hoi e.
In se tion 18.5 the veri ation of all the design spe i ations with the non-
linear system is presented. Simulations for the worst possible ombinations of
delay, mass and entre of gravity are given.
In se tion 18.6 the results of the automated evaluation pro edure are given.
The theory has been presented in hapter 3. More details, in luding pro-
gram ode for the design written in Matlab an be found in [52

18.2 Sele tion of the ontroller ar hite ture

We make use of the lassi al approa h onsisting of splitting up the ontroller
into two parts: a longitudinal and a lateral ontroller. Both of them have been
designed in two stages: the inner loop and the outer loop. The fun tion of
the inner loop is to make the air raft easy and pleasant to y, and it is often
alled a stability augmentation system. The outer loop fun tion is to repla e
the pilot for ertain ight manoeuvres su h as maintaining height and speed,
turning onto a spe ied heading, limbing at a spe ied rate, et .
Figure 18.1 shows the stru ture of the longitudinal ontroller, and gure
18.2 shows the stru ture of the lateral ontroller.

zc -
+ LonKo
-
+
-
+
T
wVc
z
VAc
-
+ LonKin
Aircraft
+
q
nx TH Actuators
nz
wV
VA

( z , q , nx , nz , wV , VA )

Figure 18.1: Longitudinal ontroller

18.2.1 Longitudinal ontroller

The hoi e of the signals to be used in a design is based on the analysis of the
system, the spe i ations and the design method.

Measurement signals
Table 18.1 shows the measurements as used by the longitudinal ontroller.

239
 vc
atan(vc/uc) c
uc
Lateral
yc - deviation
+ LatKo -
+ -
+

y
=0 LatKin
Aircraft
+

p R Actuators
r

(y,,p,r,,)

Figure 18.2: Lateral ontroller

Inner Loop ontroller

q Pit h rate
nx Horizontal load fa tor
nz Verti al load fa tor
wV z omponent of inertial velo ity
VA Air speed

Outer Loop ontroller

z z position of air raft CoG
Table 18.1: Longitudinal measurements used

Although in the inner loop only four measurements are needed to assign
four eigenvalues, two for the short period and two for the phugoid, we make
use of ve in order to prevent the slower mode of the a tuators (that of the
throttle) from be oming unstable, see [52.

Finally, the outer loop provides altitude tra king by adding a feedba k of
the altitude z.

A tuator signals

These signals are the elevator dee tion or tailplane dee tion T , and throttle
position T H .

Referen e signals
The sele tion of signals hosen as referen es has been guided by the spe i-
ations given as design riteria, see 18.3. The sele ted ones are the referen e
velo ity wV and the referen e airspeed VA for the inner loop, and the referen e
position zC for the outer loop.

240
Controller stru ture
The inner loop ontroller has a stati gain matrix a ting on the ve hosen
measured signals and on the integral of the errors of the ommanded variables
wV and VA , in the order here spe ied. The two integrators result in two addi-
tional states that must be in orporated into the linear model for the ontroller
design. The outer loop has a proportional a tion a ting on the altitude error.
No integral a tion is needed here to avoid steady state errors related to altitude
step ommands or disturban es, sin e the altitude dynami s in lude a pole at
the origin. The output of the outer loop a ts as a referen e for wV .

18.2.2 Lateral Controller

The lateral ontroller has a stru ture analogous to the longitudinal ontroller.

Measurement signals
Table 18.2 shows the measurements as used by the lateral ontroller.

Inner Loop ontroller

 Angle of sideslip
p Roll rate
r Yaw rate
 Roll angle
 Inertial tra k angle

Outer Loop ontroller

y Lateral deviation

Table 18.2: Lateral measurements used

A tuator signals
These signals are the aileron dee tion A , and rudder dee tion R .

Referen e signals
Again, the sele tion of referen es has been guided by the spe i ations given as
design riteria, see 18.3. Those sele ted are the lateral deviation for the outer
loop, and the inertial tra k angle C and sideslip angle C for the inner loop.
C is not given as a referen e signal but it is obtained from the referen e velo ity
omponents uC and vC as atan(vC =uC ). C has a null onstant referen e value
in order to keep  always lose to zero.

Controller stru ture

The inner loop ontroller has a stati gain a ting on the ve hosen signals
measured, and on the integral of the errors of the ommanded variables  and

241
, in the order spe ied here. The introdu tion of two integrators results in
two additional states that must be in orporated into the linear model for the
ontroller design. The outer loop has a proportional a tion a ting on the lateral
error. No integral a tion is required here to avoid steady state errors relating
to lateral step ommands or disturban es, sin e the lateral dynami s in lude
a pole at the origin. The output of the outer loop a ts as a referen e for the
inertial tra k angle.

18.3 Translation of design riteria into method

dependent obje tives
Out of the ve lasses of design riteria given in hapter 14, for the RCAM de-
sign, the performan e riteria are the most signi ant. These riteria are given
in terms of transient response hara teristi s to ommand signals and ross
oupling onstraints. The main hara teristi of the eigenstru ture assignment
method is that it allows the designer to satisfy spe i ations dire tly in terms
of transient response and mode de oupling. Then, these are the most suitable
riteria to be used as a guide in the design phase.
Our design is based mainly on these riteria, but we made use of linear and
non-linear losed-loop time responses and robust analysis tools as a guide in
the sele tion of the eigenstru ture.

18.3.1 Performan e Criteria

The performan e riteria an be lassied into two groups: longitudinal and
lateral. We dis uss separately for ea h group the way in whi h the spe i a-
tions have been in orporated into the design.

Longitudinal spe i ations

There are ommand response spe i ations in terms of overshoot, rise time
and settling time for three ommanded signals: airspeed, ight path angle and
altitude. These spe i ations provide a lower limit for the damping ratio and
natural frequen y of se ond order modes and for the time onstant of rst order
modes oupled with the signals.
The ight path angle is neither available as an output nor as a referen e
signal. To ope with this situation we use the relationship sin( ) = wV =V ,
where V is the total inertial velo ity. Therefore, for onstant inertial velo ity
the ight path angle may be ontrolled by means of wV . This leads us to an
interpretation of the spe i ations in terms of ommands in wV . We add a
verti al velo ity error integrator state to get good low-frequen y tra king.
We add an outer loop for tra king the referen e altitude. A signal propor-
tional to the altitude error is used as a referen e signal for wV . The proportional

242
gain is al ulated using the root lo us method in su h a way that it fulls the
transient riteria.
The spe i ation of de oupling between airspeed and altitude may be ob-
tained by de reasing ross- oupling between VA and wV .

Lateral spe i ations

There are now two ommand signal spe i ations (heading angle and lateral de-
viation) in terms of transient response hara teristi s that may be transformed
into bounded eigenvalues. The remaining spe i ations are given in terms of
behaviour in ase of engine failure and under disturban es.
The heading angle is a lateral motion state but it is neither available as an
output nor as a referen e. Instead, we make use of the inertial tra k angle that
is an output signal and is related to the heading angle by means of the equation
= + V . Although there is no referen e signal for , su h a signal may
be obtained from the referen e velo ity omponents uC and vC by means of
equation C = tan 1 (vC =uC ). An integrator is introdu ed to eliminate sideslip
errors, where the ommand signal for  is hosen as 0. Doing so, we ope with
the safety riteria of keeping  minimised at all times. An inertial tra k angle
error integrator is also introdu ed in order to avoid heading angle steady state
errors.
The lateral deviation is ontrolled in the outer loop. A signal proportional
to the lateral error is used as a referen e signal for the inertial tra k angle. The
proportional gain is al ulated using the root lo us method in su h a way that
it fulls the transient riteria.
The engine failure requirements an not be easily interpreted in terms of
eigenvalue and eigenve tor spe i ations and shall not be tested before the
phase of analysis of simulation results. However, the requirement of keeping
sideslip angle to a minimum in ase of engine failure an be translated into a
spe i ation of de oupling between the eigenvalues related with the roll motion
and those related with the lateral velo ity.

18.3.2 Robustness riteria

The eigenstru ture method is not a robust ontrol method and, although many
dierent algorithms have been proposed to enhan e the robustness, no method
will be used here. However, we make use of the stability margins given in hap-
ter 3 to measure the robustness of the feedba k system. For a hosen eigen-
stru ture the stability margins are measured and the eigenstru ture hanged
to get better stability margins. After a few steps a suitable eigenstru ture is
sele ted.

18.3.3 Other riteria

The ride quality riteria, safety riteria and ontrol a tivity riteria an not be
in orporated in an ad ho manner into the eigenstru ture method. As done

243
with robustness, they will be analysed in the ourse of the analysis of results
phase and physi al relations between their behaviour and the eigenstru ture
a hieved will be established in order to ope with them. In the sele tion of
the eigenvalues we should have in mind to assign the mode values lose to the
open-loop air raft modes to minimise the ontrol a tivity.

18.4 Design y le and ontroller derivation

The design y le is summarised in the following s heme

1. Analysis of the linear and non-linear model of the plant.

2. Sele tion of the eigenstru ture and gain al ulation.

3. Analysis of the open and losed-loop system (linear and non linear model).

4. Robusti ation pro edure (iterate).

5. Iterate 1 - 4.

As we have seen in the previous se tion, among all the design riteria only
those of the performan e riteria related to the transient response of the system
an be interpreted almost dire tly in terms of eigenstru ture. The rest of them
must be analysed after an eigenstru ture has been hosen and the ontroller
found. From this analysis another eigenstru ture will be hosen and so on.
That has been the most onsuming time task. A robusti ation pro edure or
any other form of "optimal" solution may help to break the iterative pro edure.
We will now explain the method we have used to sele t an eigenstru ture.

 We analysed the design riteria and the oupling of the modes of the
plant with the states, the inputs and the outputs. This analysis was used
as a guide in hoosing the outputs for feedba k and the eigenstru ture.
After sele ting the outputs to be used for feedba k the integrators were
added to the loop.

 We hose a set of eigenvalues that fullled the transient response limits in

a ordan e with the requirements and were lose to the natural air raft
modes. After that, their asso iated eigenve tors were hosen to get the
desired de oupling between the modes.

 With the eigenstru ture hosen, we omputed the feedba k gain and anal-
ysed:

- the stability (sin e one of the non-assigned eigenvalues might be ome

unstable or badly damped) and the stability margins

- the time response of the system

- the de oupling of the obtained modes

244
 Now the y le begins. New outputs and/or eigenstru ture must be hosen
to improve the results obtained. We have rst sele ted an eigenstru ture that
provides a eptable design riteria. We have tested dierent eigenvalues with
xed eigenve tors. On e the eigenvalues that give better stability margins have
been hosen, the eigenve tors have been hanged to try to improve robustness
and, when ne essary, de oupling. After a few steps we onvin ed ourselves that
the hosen eigenve tors ould not be improved.

18.4.1 Longitudinal ontroller

Longitudinal model
The non-linear model is used to generate linear models for ontrol law design
and to generate non-linear time histories for evaluating ontrol designs. On e
a trim ondition is established for the non-linear air raft model within the sim-
ulation environment, a linear model is generated to apture the perturbational
dynami s around the equilibrium point. The model has been linearised around
the following operating ondition: V = 80 m/s, h = 1000 m, mass = 120000
kg, gx = 0:23 and gz = 0:1. The aerodynami model is augmented with rst
order a tuator models. Two integrated error states are added to the linearised
model, one for wV and another for VA . The number of outputs is now in reased
by two and the eigenvalues for the modes of the integrators may be spe ied.
The maximum allowed transport time delay of 0.10 s is added to the model
with a rst order Pad approximation.

Inner loop ontroller design

As mentioned in hapter 3 we an spe ify as many losed-loop eigenvalues as
outputs for feedba k used. Therefore, only four measurements are needed to
spe ify the phugoid and short period modes. However, we used ve measure-
ments so that the slower mode of the a tuators (that of the throttle) is spe ied,
in order to avoid it be oming unstable.

R R
Mode Phugoid Short Period Throttle wV VA
Eigenvalues 0:4376 0:9059 0:5 2:0000 1:9000
0:0624{ 0:4388{
q xx 1x x x x
 x1 xx x x x
uB 1x 00 x 0 x
wB 00 x1 x x 0
XT xx xx x x x
XT H xx xx 1 x x
delay T xx xx x x x
delay
R T H xx xx x x x
R wV 00 xx x 1 0
VA xx 00 x 0 1

Table 18.3: Desired eigenstru ture of the longitudinal losed-loop system

245
 Table 18.3 shows the eigenstru ture hosen for the system. The state om-
ponents are given in the rst olumn, where XT is the state orresponding to
the rst order tailplane model and XT H the state orresponding to the rst
order engine model. The rst row shows the desired eigenvalues, and the de-
sired eigenve tors are shown underneath, where the symbol "x" represents the
unspe ied elements in the eigenve tors. The resulting gain is:
 
LonKin = 00::4755
0455
0:0532 0:0838 0:0169 0:0055 0:0033 0:0014
1:3063 0:3047 0:0152 0:1221 0:0004 0:0227
Loop stability margins
Figures 18.3 and 18.4 show the singular value plots of the sensitivity fun tion S,
the omplementary sensitivity fun tion T and the balan ed sensitivity fun tion
S+T at the a tuator inputs and at the sensor outputs. Tables 18.4 and 18.5
show the gain and phase margins obtained from the sensitivity fun tions.

S, T and S+T (HG) S, T and S+T (GH)

1 1
10 10

S+T S+T
0 0
10 10
magnitude

magnitude

1 S T 1 S T
10 10

2 2
10 2 1 0 1
10 2 1 0 1
10 10 10 10 10 10 10 10
frequency (rad/s) frequency (rad/s)

Figure 18.3: Longitudinal inner Figure 18.4: Longitudinal inner

loop singular values of the input loop singular values of the output
sensitivity fun tions sensitivity fun tions

Fun tion  ! Km = 1= Gain margin (dB) Phase margin (deg)

S 1.42 0.57 0.70 [-4.6,10.6 41.2
T 1.03 0.09 0.98 [-29.7,5.9 57.8
S+T 1.96 0.48 0.51 [-9.8,9.8 54.1
Table 18.4: Longitudinal inner loop stability margins at the inputs

Fun tion  ! Km = 1= Gain margin(dB) Phase margin (deg)

S 1.47 0.54 0.68 [-4.5,10.0 39.9
T 1.00 0.02 1.00 [-54.4,6.0 59.9
S+T 1.96 0.46 0.51 [-9.8,9.8 54.1
Table 18.5: Longitudinal inner loop stability margins at the outputs

Good stability margins are obtained, but we must remember that these
margins are onservative, and even better stability margins should be expe ted.

246
Outer loop ontroller design
Choosing LonKo = 0:1027, the slower roots are 0:14  0:14{, having a rise
time of 11 s, whi h is less than the spe ied rise time for an altitude ommand
(15 s), and a settling time of 35.4 s whi h is well below the required 45 s. The
gain margin for the outer loop is 13 dB at w = 0:35 rad/s and the phase margin
is 63 deg at w = 0:1 rad/s.

18.4.2 Lateral ontroller

Lateral model
The pro edure followed in dealing with the lateral model is analogous to the
one used with the longitudinal model.
A linear model is generated from the same trimmed ondition used to obtain
the linearised longitudinal model. The a tuator dynami s have been added to
the linear model by augmenting it with orresponding states. Also, we added
the two integrated error states and a 0.10 s delay with a rst order Pad ap-
proximation.

Inner loop ontroller design

Table 18.6 shows the eigenstru ture hosen. The state omponents are given in
the rst olumn, where XA is the state orresponding to the rst order aileron
dee tion model and XR the state orresponding to the rst order rudder
dee tion model.

R R
Mode Dut h roll Spiral Roll Sub. Heading  
Eigenvalues 0:8  0:6{ -0.4 -1 -0.8 -1.5 -1.4
p 00 x 1 x 0 x
r 1x x 0 x x x
 00 1 x x 0 x
xx x x 1 x x
vB x1 0 0 0 x 0
XA xx x x x x x
XR xx x x x x x
Delay a xx x x x x x
R r
Delay xx x x x x x
R xx x x x 1 x
 xx x x x x 1

Table 18.6: Desired eigenstru ture of the lateral losed-loop system

The resulting gain is:

 
LatKin = 3:6246 1:7016 2:9057 3:0480 13:1933 0:6869 2:2288
1:5216 0:0782 2:4251 0:2268 1:0320 0:7237 0:1820

247
 Loop stability margins
The results obtained at the a tuator inputs and at the sensor outputs are
summarised in Table 18.7 and Table 18.8, respe tively. Figures 18.5 and 18.6
show the singular values plots of the sensitivity fun tions.

Fun tion  ! Km = 1= Gain margin (dB) Phase margin (deg)

S 1.43 0.65 0.70 [-4.6,10.4 40.8
T 1.18 0.21 0.85 [-16.2,5.3 50.0
S+T 2.19 0.46 0.46 [-8.6,8.6 49.2
Table 18.7: Lateral inner loop stability margins at the inputs

Fun tion  ! Km = 1= Gain margin(dB) Phase margin (deg)

S 1.71 0.50 0.59 [-4.0,7.7 34.1
T 1.01 0.04 0.99 [-43.3,6.0 59.5
S+T 2.19 0.46 0.46 [-8.6,8.6 49.2
Table 18.8: Lateral inner loop stability margins at the outputs

S, T and S+T (HG) S, T and S+T (GH)

1 1
10 10

S+T S+T
0 0
10 10
magnitude

magnitude

1 S T 1 S T
10 10

2 2
10 2 1 0 1
10 2 1 0 1
10 10 10 10 10 10 10 10
frequency (rad/s) frequency (rad/s)

Figure 18.5: Lateral inner loop sin- Figure 18.6: Lateral inner loop sin-
gular values of the input sensitivity gular values of the output sensitiv-
fun tions ity fun tions

Outer loop ontroller design

Choosing LatKo = 0:001, the gain margin for the outer loop is 13.9 dB at
w = 0:29 rad/s and the phase margin is 63.9 deg at w = 0:08 rad/s. The
slower roots are 0:11  0:10{ that have a rise time of 15 s and a setting
time of 44 s. Therefore, the transient spe i ations for a lateral deviation step
ommand are met.

248
 18.5 Analysis of the resulting ontroller
The ontroller was designed by an iterative method onsidering riteria of per-
forman e and robustness in the linear model, without taking into a ount non
linearities. In this se tion we present the veri ation of all the design spe i-
 ations with the non-linear system. The des ription of the design riteria is
given in hapter 14.
All the simulations are run using the following onguration (see 18.4):
mass = 120.000 kg,
x = 0:23 ,
y = 0 and
z = 0:1 ; and the initial
onditions are: VA = 80 m/s, altitude = 1000 m, = 90 deg, = 0 deg and
 = 1:65 deg.

18.5.1 Non-linear simulation observing performan e ri-

teria
Figure 18.7 shows the response to lateral deviation, altitude, heading angle,
ight path angle and airspeed ommands. It an be seen that all performan e
riteria on erning altitude, heading angle, inertial ight path angle, airspeed
and ross oupling from airspeed to altitude are fullled. In the lateral deviation
step ommand a step hange of 100 m has an overshoot of 13 %, however in the
linearised model the overshoot is smaller than the spe ied 5 %, but the lateral
deviation is redu ed to 10 % in less than 30 s, a ording to spe i ations.
Psi_dot (deg/sec)
Step of 100 m
Lat. desv. (m)

85
Psi (deg)

100 2
50 90 0
2
0 95
0 10 20 30 40 0 50 100 150 200 0 50 100 150 200

Step of 30 m in altitude Step of 30 m in altitude

Altitude (m)

80.5 5 10
Beta (deg)
VA (m/s)

Phi (deg)

1020
80 0 0
1000
79.5 5 10
0 10 20 30 40 0 10 20 30 40 0 50 100 150 200 0 50 100 150 200

Step of 3 deg in Psi Step of 3 deg in Gamma

Gamma (deg)

81 0.2
V_A (m/s)
Psi (deg)

ny (g)

88 2 80 0
90 0 79 0.2
0 10 20 30 40 0 10 20 30 40 0 50 100 150 200 0 50 100 150 200

Step of 13 m/s in VA Step of 13 m/s in VA

Altitude (m)

Alpha (deg)

Altitude (m)

1010 2 1005
VA (m/s)

90
1000 1.5 1000
85
80 990 1 995
0 10 20 30 40 0 10 20 30 40 0 50 100 150 200 0 50 100 150 200
time (sec) time (sec) Time (sec) Time (sec)

Figure 18.7: Response of the non- Figure 18.8: Response of the non-
linear model to ommand signals linear model to an engine failure

18.5.2 Non-linear simulation of an engine failure

The results are given in Figure 18.8. We an see that:  is less than the spe ied
10 deg, its steady state deviation does not ex eed 5 deg and, when the engine
is restarted,  is redu ed to zero with an overshoot that sligtly ex eeds the
spe i ation of 50 %;  is qui kly minimised; and _ is well below the spe ied
3 deg/s.
The omfort and safety riteria are also veried: the airspeed is always
higher than 79 m/s, whi h is mu h bigger than the spe ied 62:2 m/s =

249
1:2Vstall ; the angle of atta k   1:7 deg is less than the spe ied 12 deg;
and the lateral a eleration ny in reases up to 0:1 g, that is less than the
spe ied 0:2 g.
Finally, the variations in altitude are less than 2 m.

18.5.3 Non-linear simulation observing ride quality and

safety riteria
The spe i ations for ride quality riteria are a maximum verti al a eleration
and a maximum lateral a eleration. Figure 18.9 shows the lateral a eleration
90 deg turn and the verti al a eleration in a hange from = 0 deg to
in a
= 6 deg and from -6 deg to -3 deg. These values orrespond to segments
II and III of the Automated Evaluation Pro edure (18.6), respe tively. It an
be seen that lateral a eleration ny is always within the limits and the verti al
a eleration nz slightly surpasses the maximum level.

90 degrees turn gamma rotation to 6 and 3 degree

0.02 0.05
ny (g)

nz (g)

0 0

0.02 0.05
0 20 40 60 80 0 50 100
time (sec) time (sec)

Figure 18.9: Ride quality riteria observation

Figure 18.10 depi ts the results of the non-linear simulation showing the
safety riteria. The simulations orrespond to the landing approa h of the
evaluation pro edure as used in 18.6, whi h is a good representation of all
possible ight onditions.
Alpha (deg)
V_A (m/s)

90
10
80
70 0
0 200 400 0 200 400
10
Beta (deg)

20
Phi (deg)

0 0
20
10
0 200 400 0 200 400
time (sec)

Figure 18.10: Safety riteria observation

The safety riteria are fullled sin e the airspeed is always well above
1:2Vstall = 62:2 m/s; the maximum angle of atta k  remains within the limits,
its maximum value is observed during the turn, but is well below the limit of
12 deg; the roll angle  remains also within the limits, but in the turn is near
to the limit of 30 deg; and the sideslip angle  is qui kly minimised at all times.

250
 Detailed response to an engine failure for airspeed and angle of atta k has
been given in the previous se tion.

18.5.4 Non-linear simulation under moderate turbulen e

onditions
We will give an analysis of the non-linear systems under moderate turbulen e
onditions. The spe tra of the turbulen e we used in the simulations are de-
s ribed in hapter 14 (14.2.6).
Figure 18.11 shows a tuator behaviour. All the requirements on the mean
of a tuators are fullled. Moreover, all RMS values are less than the limits (the
spe ied values are given in parentheses):

- The mean aileron rate is 0.42 deg/s ( < 8:25 deg/s) with a RMS of 7.30
deg/s

- The mean tailplane rate is 0.12 deg/s ( <5 deg/s) with a RMS of 1.03
deg/s

- The mean rudder rate is 0.08 deg/s ( < 8:25 deg/s) with a RMS of 1.97
deg/s

- The mean throttle rate is 0.09 deg/s ( < 0:24 deg/s) with a RMS of 1.50
deg/s
dt_dot (deg/sec) da_dot (deg/sec)

Open loop Close loop

5 5
Phi (deg)

Phi (deg)

20
da (deg)

20
0 0 0 0
20 20
5 5
0 50 100 0 50 100 0 50 100 0 50 100

5 5
Beta (deg)

Beta (deg)

20
dt (deg)

20
0 0 0 0
20 20
5 5
0 50 100 0 50 100 0 50 100 0 50 100
dr_dor (deg/sec)

85 85
Psi (deg)

Psi (deg)
dr (deg)

20 20
0 0 90 90
20 20
95 95
0 50 100 0 50 100 0 50 100 0 50 100
dth_dot (deg/sec)

10 5 85 85
dth1 (deg)

Chi (deg)

Chi (deg)

5 0 90 90

0 5 95 95
0 50 100 0 50 100 0 50 100 0 50 100
time (sec) time (sec) time (sec) time (sec)

Figure 18.11: Control a tivity un- Figure 18.12: Responses under

der moderate turbulen e onditions moderate turbulen e onditions of
angles ,  , and 
Figure 18.12 shows in open and losed-loop, the roll angle , the sideslip
angle  , the heading angle and the inertial tra k angle  under these moderate
turbulen e onditions. In losed-loop,  always remains within the spe ied
limit of 5 deg. Table 18.9 shows the RMS values of the errors in open and losed-
loop. The losed-loop values of ,  and  are less than their orresponding
open-loop values, however the RMS of is higher in losed-loop than in open-
loop.

251
   
Open loop RMS (deg) 2.04 1.06 0.89 0.97
Closed loop RMS (deg) 0.65 0.96 1.45 0.28

Table 18.9: RMS of the errors in angles ,  , and  in open and losed-loop

18.5.5 Non-linear simulations observing robustness rite-

ria
Figures 18.13-18.15 show the simulation results obtained at the nominal, min-
imum and maximum values of time delay, mass and entre of gravity, all ases
at the design airspeed of 80 m/s, giving a set of 81 simulations. In the plots,
the angles are given in radians, the displa ements in meters and the velo ities
in m/s.
Figure 18.13 shows the lateral step response and the altitude step response.
We an see that there is very little ontrol a tivity, there are almost no over-
shoots and the settling times are well below the spe ied values.
Figure 18.14 shows the airspeed-altitude ross- oupling. We an see that
the deviations are always within the limits.
Figure 18.15 shows the roll and heading angles at engine failure. In all ases
the performan e riteria are fullled: the roll angle never ex eeds 10 deg and
its steady state deviation does not ex eed 5 deg; the heading rate is always well
below the spe ied 3 deg/s, and the sideslip angle is qui kly minimised.
In [52, the simulation results an been found for the inner loop response to
heading and ight path angle steps ommands, and for air speed step and wind
step responses. With respe t to the inner loop response to heading and ight
path angle steps, in all ases the overshoot is lower than the spe ied 5 % and
the settling time is less than the spe i ations. Regarding the airspeed step
and wind step response, all the performan e riteria are fullled in the airspeed
ommand. The response to a wind step has three dierent hara teristi s whi h
orrespond to the three dierent values of the mass whi h have been onsidered.
In ase of minimum and maximum mass values, the deviation in airspeed is
larger than 2.6 m/s after 15 s of the step, and in ase of nominal mass value
the spe i ation is fullled.
In [52, the simulation results obtained for all possible worst onditions of
time delay, mass and entre of gravity, at a speed of 63.7 m/s and at maximal
ap speed (90 m/s) an also be found.
The worst results are obtained in the engine failure at 63.7 m/s, with min-
imum mass and with
x = 0:31 . Consequently, a new set of ontroller gains
was omputed using the linearised model at 63.7 m/s velo ity, but without
hanging the rest of the trimming parameters, and with the same eigenstru -
ture as before. The resulting ontroller and simulation results, obtained by
making use of this ontroller when simulating at 63.7 m/s and for the worst
ondition of time delay, mass and entre of gravity, are given in [52.
Similar performan e and robustness results to those obtained with the on-
troller designed at the original speed of 80 m/s are obtained, thus gain s hedul-

252
 Lateral step response Altitude step response
y y_c . z z_c .
1.5 1.5

1 1

0.5 0.5

0 0

0.5 0.5
0 10 20 30 40 0 10 20 30 40
time [s] time [s]
vv vv_c . wv wv_c .
1 1

0.5 0.5

0 0

0.5 0.5
0 10 20 30 40 0 10 20 30 40
time [s] time [s]
da dt
0.1
0.02
0.05 0.01

0 0

0.01
0.05
0.02
0.1
0 10 20 30 40 0 10 20 30 40
time [s] time [s]
dr throttle L R
0.05 0.01

0.005

0 0

0.005

0.05 0.01
0 10 20 30 40 0 10 20 30 40
time [s] time [s]
ident: td012:m012:x012:z012:ex7

Figure 18.13: Lateral and altitude step response at the design airspeed

Airspeed Altitude cross coupling

z z_c . z (va_c)
20
0
10

20 0

10
40
20
0 10 20 30 40 0 10 20 30 40
time [s] time [s]
va (z_c) va va_c .
2

15
1
10
0
5
1
0

2 5
0 10 20 30 40 0 10 20 30 40
time [s] time [s]
dt (z_c) dt (va_c)
0.2 0.2

0.1 0.1

0 0

0.1 0.1

0.2 0.2
0 10 20 30 40 0 10 20 30 40
time [s] time [s]
throttle L R throttle L R
0.2 0.2

0.1 0.1

0 0

0.1 0.1

0.2 0.2
0 10 20 30 40 0 10 20 30 40
time [s] time [s]
ident: td012:m012:x012:z012:ex7

Figure 18.14: Airspeed-altitude ross oupling at the design airspeed

253
 Roll angle and heading response to right engine failure
phi R e_f . psi R e_f .
1
0.2
0.5
0.1
0 0
0.1
0.5
0.2
1
0 10 20 30 40 0 10 20 30 40
time [s] time [s]
beta R e_f . psid R e_f .
0.1
0.2
0.05
0.1
0 0
0.1
0.05
0.2
0.1
0 10 20 30 40 0 10 20 30 40
time [s] time [s]
da R e_f . dt R e_f .
0.1
0.5
0.05

0 0

0.05
0.5
0.1
0 10 20 30 40 0 10 20 30 40
time [s] time [s]
dr R e_f . throttle L R
0.2
0.5
0.1

0 0

0.1
0.5
0.2
0 10 20 30 40 0 10 20 30 40
time [s] time [s]
ident: td012:m012:x012:z012:ex7

Figure 18.15: Roll angle and heading response to right engine failure at the
design airspeed

ing with respe t to velo ity should be used.

18.6 Results of the automated evaluation pro e-

dure
This se tion presents the results of the evaluation pro edure, as proposed in
14.3.3 onsisting of a landing approa h.

Segment I
Figure 18.16 shows the performan e of the ontroller in this segment, and it
an be seen that the lateral deviation is always less than 20 m. Therefore, the
ontroller omplies with the orresponding spe i ation. Moreover, the four
plots are almost the same, whi h means small sensitivity to time delay and to
horizontal entre of gravity variations.

Segment II
Figure 18.17 gives the behaviour of the model in this manoeuvre. It an be seen
that the traje tory of the model surpasses the bounds marked in the plots but
the lateral deviation never ex eeds the maximum value of 200 m and at the end
the lateral deviation is lose to zero. The lateral a eleration never surpasses

254
 the maximum allowable value (see Figure 18.9). Moreover, the model has a
very smooth turn, fulls all the performan e design riteria and our attempts
to have a traje tory within the bounds diminished the stability margins, so we
a epted it as is.

Segment III
Figure 18.18 represents the behaviour of the model in the des ent phase. It
an be seen that the traje tories of the model surpass the bounds marked in
the plots although the verti al deviation never ex eeds the maximum value of
20 m and at the end of the segment the deviation is lose to zero. In Figure
18.10 we an see that the speed variation is well below the allowed 4 m/s.
Moreover, the model has a very smooth transition during the entire segment,
although the verti al a eleration is a little bit high at some points. We an
see in Figure 18.9 that the verti al a eleration slightly surpasses the maximum
allowed value. This is ree ted in the omfort index in Table 18.10. Sin e the
rest of the design riteria are fullled and our attempt to diminish this value
produ es worse results, we a epted it.

Segment IV
Figure 18.19 shows the behaviour of the model in this segment. It an be seen
how the traje tories fall inside the bounds during the entire segment. The rest
of the spe i ations are fullled by the ontroller.

Numeri al results
Table 18.10 summarises the results as obtained by the ontroller along the
landing approa h. For full details see hapter 14. In general the results are
good, ex ept for the omfort riterion in Segment III. The problem with the
omfort has already been explained. It is basi ally due to a small high level of
the verti al a eleration.

First segment: top view Second segment: lateral deviations

300

100
200

50
lateral deviation [m]

100
xdeviation [m]

0 0
0 a b 1
1 c d 2

100
50

200
100

300
20 18 16 14 12 10 8 6 4 2 0 1 2 3 4 5 6 7 8
yposition (YE) [km] along track distance from point 1 [km]

Figure 18.16: Segment I: The ee t Figure 18.17: Segment II: Lateral
of engine failure with bounds deviations during the 90 degrees
turn with bounds

255
 Third segment: altitude deviations Fourth segment: altitude deviations
30 30

20 20

altitude deviation [m]

altitude deviation [m]

3
10 2 f 3 10
4
g h
0 0

10 e 10

20 20

30 30
16 15 14 13 12 11 11 10 9 8 7 6 5 4 3 2 1
xposition (XE) [km] xposition (XE) [km]

Figure 18.18: Segment III: Verti al Figure 18.19: Segment IV: Verti-
deviations during the -6 and 3 de- al deviations during the nal ap-
grees glidslope with bounds proa h with bounds

Segm. I Segm. II Segm. III Segm. IV Total

Performan e 0.0764 0.4964 0.3285 0.1905 0.2730
Perf. Dev. 0.0309 0.0161 0.4926 0.1873 0.1817
Comfort 0.5432 0.7340 1.1808 0.4674 0.7314
Safety 0.0038 0.0382 0.0070 0.0345 0.0209
Power 0.0037 0.0027 0.0150 0.0309 0.0131

Table 18.10: Numeri al results of the evaluation pro edure

18.7 Con lusions

In this hapter the eigenstru ture method has been applied to the RCAM
ben hmark problem.
The design was ompleted by making use of the lassi al approa h. This
onsists of splitting up the ontroller into two parts, a longitudinal and a lateral
ontroller, and in using a standard inner-outer loop ontrol stru ture. In every
inner loop the feedba k outputs, ommand signals and integrated outputs have
been hosen guided by the design spe i ations. For both inner loops a onstant
gain feedba k matrix has been al ulated using the eigenstru ture te hnique.
This method allows the designer to satisfy dire tly performan e riteria given
in terms of damping, settling time and de oupling, but not to ope dire tly with
system un ertainties. However, the eigenstru ture was hosen in an iterative
way, so that good stability margins were obtained. In both outer loops a
onstant s alar gain has been used.
When analysing the ontroller with the non-linear model (18.5) all the
design riteria are fullled, but the maximum verti al a eleration is surpassed
in ertain onditions and the RMS of the heading angle error in losed-loop is
greater than in open-loop.
Good robustness results are obtained with respe t to variations of the mass,
entre of gravity and transport time delay. However, gain s heduling should
be used with respe t to velo ity sin e bad results are obtained with speed

256
variations, mainly for the engine failure ase, as explained in 18.5.
The results obtained in the automati evaluation pro edure demonstrate
good good fulllment of all of the design riteria ex ept the omfort riteria.
In our design this is due to the fa t that the verti al a eleration annot be
diminished without violating other performan e riteria.
The sele tion of a good eigenstru ture follows an iterative pro ess that an be
time onsuming. The pro ess ould be shortened if some optimisation method
were used. Without any optimisation pro ess, doubts about how good the
ontroller is will always remain.

257
19. A Modal Multi-Model Approa h

Carsten Dll , Jean-Franois1 Magni1

1

and Yann Le Gorre

Abstra t. This paper presents a modal multi-model approa h ap-
plied to the RCAM design hallenge. Inner feedba k loops (sta-
bilization) are designed by eigenstru ture assignment. Equations
orresponding to de oupling properties for a representative set of
linearized models are olle ted. In the lateral ase, the number of
equations is small enough so that these equations an be solved us-
ing a proportional feedba k. In the longitudinal ase, the use of a
dynami ontroller turns out to be ne essary. Outer loops (tra k-
ing) are designed separately using lassi al ontrol.

19.1 Introdu tion

The RCAM design problem is aimed at keeping as lose as possible to industrial
requirements. Espe ially, proportional or dynami gains with minimum number
of states are required. Under this onstraint we tried to redu e the usual
s heduling as mu h as possible by improving robustness. This task was even
more di ult on a ount of the unusual nature of the RCAM problem, having
its own evaluation riteria whi h diered somewhat from usual industrial ones.
The main dieren e between a tual industrial problems and RCAM prob-
lem omes from the fa t that the autopilot we are looking for is the same
for ruising, beam apture and landing. It would have been possible to de-
sign three stru tures and to onsider ommutations, but this kind of problem
is well understood in industry, therefore no improvement ould be expe ted
from our ontribution. So, we preferred to design a single stru ture and on-
sequently, the autopilot design we propose has faster than a typi al autopilot
(indu ing some overshoot and omfort deterioration). Nevertheless we believe
that the methodologies proposed to redu e s heduling are still valid (maybe
more e ient) in a more realisti industrial setting.
Se tion 19.2 details the ontroller stru ture used and the hosen referen e
inputs and measurements. Then, the spe i ations are translated into method
dependent obje tives in Se tion 19.3. In Se tion 19.4 the proposed ontroller
is derived. An analysis of robustness and performan es is proposed in Se tion
19.5.

1
CERT ONERA, Dpartement d'tudes et Re her hes en Automatique, BP 4025, F31055
Toulouse Cedex, Fran e.

258
Outer loop design. A simple preliminary inner loop is used for this purpose. It
appeared that it is ne essary to use a sophisti ated outer loop with feedforward.
Indeed, the use of feeforward permits a designer to meet the design spe i a-
tions with a mu h slower inner loop; in turn, most of the inner loop degrees of
freedom an be devoted to robustness and performan e. The proposed outer
loop turned out to be so robust that it was not ne essary to update it when
several kinds of inner loops were being tested.

Inner loop design. First, standard eigenstru ture assignment was used. Per-
forman e is easily met by this approa h but it is more di ult to ensure ro-
bustness. For the lateral hannel it is shown that the initial eigenstru ture
assignment an be improved, without gain s heduling, so that robustness re-
quirements ould be met. For the longitudinal hannel, we did not nd a satis-
fa tory proportional gain. A non-s heduled low dimensional dynami feedba k
is proposed.
The analysis of robustness is made by onsidering a set of linearized models
(see page 265). In order to assess damping ratio and settling time, the poles
are plotted; for ross- oupling, the step responses of all the linearized models
are plotted.
More details on erning the results an be found in [55. The theory is
briey presented in Chapter 3 and more details an be found in [150.

19.2 Sele tion of the Controller Ar hite ture

We shall onsider inner and outer loops. Generally speaking,

 the inner loop stabilizes and augment the handling qualities as well as
providing robustness

 the outer loops guide the air raft along a given traje tory.

A traditional approa h to the ontroller ar hite ture is to split the ontroller

into two omponents: longitudinal and lateral ontrollers. These two parts are
treated separately from one another, ex ept for the turn ompensation (denoted
q oord).
 2
longitudinal ontroller : tra king of the total inertial speed V and of the
altitude z. The inner loop onsists of a non-intera tive ontroller, with
integral ee t, relative to the verti al speed wV and the inertial speed V.
 lateral ontroller: tra king of the sideslip angle  and of the lateral de-
viation yB . The inner loop onsists of a non-intera tive ontroller, with
integral ee t, relative to the sideslip angle  and roll angle .
2
We dis overed too late that we implemented tra king of the inertial speed instead of
tra king of the airspeed. Results on erning tra king of the airspeed an be found in [55.
Robustness issues are similar in both ases. The main dieren e is that, on a ount of the
ee t of turbulen es on ontrol surfa e a tivity, the bandwidth of the longitudinal ontroller
proposed in this hapter needs to be redu ed.

259
Longitudinal measurements (ordering is onsistent with the feedba k gains given
later). Velo ity in vehi le-axis z dire tion (wV ), total velo ity (V ), pit h rate
q
( ), verti al load fa tor (nz ), integrator over wV wV; , integrator over V V
and for the outer loop: position of CoG in earth-xed frame (z ). See also
Figure 19.1.

 p r
Lateral measurements. Angle of sideslip ( ), roll rate ( ), yaw rate ( ), roll
angle (), integrator over   , integrator over   and for the outer loop:
ight path heading angle (), lateral deviation (eyb ), integrator over eyb . See
also Figure 19.2.

The used referen e signals are given in Figures 19.1 and 19.2. The turn om-
pensation is
g sin2 g
q oord =
V os g
The generation of g and g omes from the well known relations (in whi h
V = 0):  
V _
 = + V ;  = atan
g
The longitudinal ontroller is shown in Figure 19.1 and the lateral one
in Figure 19.2. Note that the subs ript g indi ates signals oming from the
traje tory generator.

wV;g

- - Ko ?wV;
zg - - RR -- -wV-R z
6 Vg --6 --- Ki T--
Air raft
V-
6 q-
delays

q oord - -- nz-
a tuators
T H
6
6

Figure 19.1: Longitudinal autopilot stru ture. Ki is a non-s heduled 4th order
transfer matrix.

19.3 Translation of Design Criteria into Method

Dependent Obje tives
19.3.1 Settling time and overshoot riteria
These riteria are easily taken into a ount by appropriate hoi e of eigenvalues
(real part to ensure settling time, damping ratio to redu e overshoot). The

260
g
g R y
yB;g -= 0- R -
--K ?- -K ? -- RR -- -

Air raft
R 
6 6   0 --6 --- K -
. o2 o1
.g
A delays 
6 - i R
a tuat.

p
6 r

_ g
- atan ( Vg
) -
g
- - ..... -
R g -
g
(0)
6 6
( V  0)

Figure 19.2: Lateral autopilot stru ture with a referen e signal generator based
on _ . Ki is a non-s heduled proportional feedba k.

hoi e of eigenvalues in the longitudinal hannel ase is made using the Naslin
rule (see [30).
Note that the use of feedforward terms interferes quite a lot with the hoi e
of eigenvalues. The onstraints on eigenvalues that we use were derived by trial
and error (see Figure 19.1 and 19.2) after feedforward terms were introdu ed
into the outer loops.

Related design riteria. Altitude response, ight path angle response, inertial
speed response, heading angle response and overshoot limitations (passenger
omfort riteria).

19.3.2 De oupling riteria

These riteria are taken into a ount by appropriate hoi e of eigenve tors (see
Chapter 3). De oupling between V and h is taken into a ount by redu ing
ross- oupling between V and wV .
Related design riteria. Cross oupling between V and h, ross- oupling be-
tween  and .

19.3.3 Tra king riteria

The stru ture (integrators) of the ontrol law permits the designer to take these
riteria into a ount. In the inner loop, an integrator is put in series with the

261
dieren e between a measurement ( V , wB ,  and ) and the orresponding
input referen e (see Figures 19.1 and 19.2). In the outer loop, a simple gain
su es to ensure zero steady-state tra king error, provided that there is at
least one open-loop integrator. In Figures 19.1 and 19.2, these integrators
are shown expli itly (integrators between wV and z and between  and y).
Related design riteria. Speed response, lateral deviation, heading angle re-
sponse and roll angle response in the ase of engine failure.

19.3.4 Robustness riteria

Robustness is treated by using a multi-model approa h as presented in Chapter
3. This approa h permits us to treat performan e robustness together with sta-
bility robustness. The hanges in stability and performan e due to parameter
variations should be a eptable, for position of the enter of gravity, mass and
delays.

19.3.5 Other riteria

The design riteria whi h remain here on ern maximum values of several sig-
nals (asso iated with safety riteria) and RMS values. To take them into a -
ount we just  rossed the ngers. In fa t, these riteria give limitations to
the fastness of our ontrol laws, while other riteria are more easily met when
dynami s are a elerated. They were he ked during simulation runs.

Related design riteria. Maximum verti al a eleration, maximum lateral a -

eleration, angle of atta k, airspeed, roll angle. Under moderate turbulen e
onditions mean a tuator rates of aileron, tailplane and rudder and mean throt-
tle.

19.4 Design Cy le and Controller Derivation

Briey, the design y le is divided into two steps. The rst step (19.4.1,19.4.2)
on erns the outer loop design, for whi h a preliminary inner feedba k relative
to the nominal model is designed. Then, the se ond step (19.4.4,19.4.5) on-
sists of nding a robust ontroller whi h has similar properties all over the
ight domain. The rst step is an iterative pro edure involving:

 tuning of the eigenvalues assigned by the preliminary inner loop,

 outer loop gain omputation,

 hoi e of feedforward signals in the stru ture (Figures 19.1, 19.2).

The iterations ended when all ex ept robustness riteria were met at the on-
sidered nominal operating point. The details an be found in Chapter 5 of [55.
Here we prefer to present more pre isely the se ond step: robust inner loop
derivation, as it is the most innovative part of our ontribution.

262
19.4.1 Outer longitudinal loop design
First, a preliminary inner ontrol law is omputed. For that we use a model
3
(nominal) in the middle of the ight domain . Four eigenve tors orresponding
to eigenvalues f 0:99j 0:90; 0:63j 0:57g are assigned su h that C1 vi = 0 i.e.
these modes are de oupled from wV (= C1 x). Two eigenvalues f 0:14  j 0:15g
are assigned su h that C2 vi = 0 i.e. this mode is de oupled from V (= C2 x).
These eigenvalues are the roots of a Naslin polynomial. The resulting gain is:

 
Ki;lon = 0:0217 0:0026 0:451 0:116 0:0063 0:0007 (19.1)
0:0036 0:0113 0:342 0:072 0:0009 0:0013
In 19.4.4 a robust ontrol will be derived. It must orrespond to a behaviour
similar to the one obtained using Ki;lon over all the ight envelope, so that the
outer loop omputed now will not need to be updated.
With the above initial inner loop, the outer loop is designed as follows.
Considering settling time, the outer loop must be designed in su h a way that
all poles have a real part about 0:35. Using the root lo us approa h, three
eigenvalues are assigned on the same verti al line 0:325. The orresponding
gain is:
Ko;lon = 0:124 (19.2)

19.4.2 Outer lateral loop design

As for the longitudinal hannel, a preliminary inner loop is needed. The eigen-
values assigned here were obtained by trial and error within a design loop in-
luding outer loop feasibility tests and feedforward signal onsiderations. Three
eigenve tors orresponding to eigenvaluesf 0:7j 0:7; 0:62g are assigned su h
that C4 vi = 0 i.e. these modes are de oupled from (= C4 x). Three eigenval-
ues f 1:7; 1:48; 1:30g are assigned su h that C1 vi = 0 i.e. these modes are
de oupled from  (= C1 x). The resulting gain is:

 
2:68 1:98 1:86 3:58 0:30 1:51
Ki;lat = (19.3)
1:64 0:049 2:38 0:16 0:93 0:14
With the above initial inner loop, the outer loop is designed in two steps.
First outer loop (see Fig. 19.2). Considering settling time onstraints, Ko1;lat
must be su h that no pole has real part larger than 0:3. Using the root lo us
approa h, it was possible to pla e the slowest eigenvalue at 7 = 0:3 + 0:2 j .
This hoi e also satises damping ratio requirements. The orresponding gain
is:
Ko1;lat = 1:41 (19.4)

Se ond outer loop (see Fig. 19.2). Outer loop 2 onsists of a single-input multi-
output system. Measurements are the lateral deviation eyb and its integral

3
Mass m = 120000 kg , airspeed VA = 80 m/s, horizontal CoG lo ation gx 
= 0.23 ,
verti al CoG lo ation gz  and a time delay 0.075 s.
= 0.0

263
value; the input is the aileron ontrol signal. Two poles are assigned using the
te hnique of Chapter 3. Finally:

Ko2;lat = [ 0:0012 0:000028 (19.5)

19.4.3 General omments on inner loop design

Considering the initial inner loop design given in 19.4.1 and 19.4.2, in view of
the results shown Figures 19.3 (left), 19.4 (top) and 19.5 (top-left) the following
omments hold:

 Robustness is not satisfa tory.

 All performan es are good.

Therefore, it remains to design ontrol laws whi h lead to the losed-loop be-
haviour of the initial step, but all over the ight domain.
Robustness will be assessed by onsidering a bank of models overing, more
or less, the entire ight envelope. This bank of models is dened in Table 19.1.
We shall onsider pole maps and step responses showing ross oupling for all
these models. For a demand of V of 13 m/s, wV must be smaller than 0.7 m/s,
for a demand of wv
V must be smaller than 1 m/s. For a demand
of 4.2 m/s,
o o o o
of  of 20 ,  must remain smaller than 1 and  < 1 while  = 1 .

19.4.4 Inner longitudinal loop design

Closed loop poles: longitudinal Closed loop poles: longitudinal with dynamic gain

4.5 4.5

4 4

3.5 3.5

3 3

2.5 2.5

2 2

1.5 1.5

1 1

0.5 0.5

0 0

0.5 0.5
5 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0 5 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0

Figure 19.3: Longitudinal losed-loop poles. Left: initial design, right: pro-
posed 4th order dynami gain

In order to have an e ient multi-model based approa h, ompatibility

between design obje tives from one model to one other must be ensured. For
instan e, we annot expe t to assign exa tly the poles of two models whi h are
very lose. It is suggested:

264
 Model des ription
Number Mass hor. CoG vert. CoG Airspeed Delay
m [t gx [ gz [ V [m=s Delay [s
0 120 0.23 0 80 0.075

1 150 0.15 0 80 0.075

2 150 0.15 0.21 80 0.075
3 150 0.31 0 80 0.075
4 150 0.31 0.21 80 0.075

5 100 0.15 0 80 0.075

6 100 0.15 0.21 80 0.075
7 100 0.31 0 80 0.075
8 100 0.31 0.21 80 0.075

9 150 0.15 0 70 0.075

10 150 0.15 0.21 70 0.075
11 150 0.31 0 70 0.075
12 150 0.31 0.21 70 0.075

13 150 0.15 0 90 0.075

14 150 0.15 0.21 90 0.075
15 150 0.31 0 90 0.075
16 150 0.31 0.21 90 0.075

17 120 0.15 0 80 0.05

18 120 0.31 0 80 0.05
19 120 0.23 0 80 0.05
20 120 0.23 0 80 0.1

21 100 0.15 0 60 0.075

22 100 0.15 0.21 60 0.075
23 100 0.31 0 60 0.075
24 100 0.31 0.21 60 0.075

25 100 0.15 0 90 0.075

26 100 0.15 0.21 90 0.075
27 100 0.31 0 90 0.075
28 100 0.31 0.21 90 0.075

29 120 0.23 0 60 0.075

30 120 0.23 0 65 0.075
31 120 0.23 0 70 0.075
32 120 0.23 0 75 0.075
33 120 0.23 0 85 0.075
34 120 0.23 0 90 0.075

35 150 0.31 0.21 75 0.075

36 150 0.31 0.21 85 0.075

Table 19.1: Considered linearized models

265
 W_V versus demand of V: 13 m/s W_V versus demand of W_V: 4.2 m/s
1 6

0.5 4
W_V : m/s

W_V : m/s
0 2

0.5 0

1 2
0 20 40 60 0 20 40 60

V versus demand of V: 13 m/s V versus demand of W_V: 4.2 m/s

15 1

0.5
10
V : m/s

V : m/s
0

5
0.5

0 1
0 20 40 60 0 20 40 60
t:s t:s
W_V versus demand of V: 13 m/s W_V versus demand of W_V: 4.2 m/s
0.2 5

4
0
3
W_V : m/s

W_V : m/s

0.2 2

1
0.4
0

0.6 1
0 20 40 60 0 20 40 60

V versus demand of V: 13 m/s V versus demand of W_V: 4.2 m/s

15 0.4

0.2
10
0
V : m/s

V : m/s

0.2
5
0.4

0 0.6
0 20 40 60 0 20 40 60
t:s t:s

Figure 19.4: Longitudinal step responses. Top: initial design, bottom: pro-
posed 4th order dynami gain

266
  to onsider models as far as possible from ea h other (for example, models
12 and 25 in Table 19.1) so that ompatibility problems be ome less
stringent.

 to assign eigenve tors by means of proje tions, to ensure ontinuity of

requirements.

The design y le redu ed to rstly ompute a new initial gain relative to

model 12. Then, analysis of all models is done onsidering this initial feedba k.
Poorly damped high frequen y poles arise. They belong to models around
model number 25 (low mass, high speed). For robustness improvements the
poles of model 25 are shifted towards the left, while some poles of the low
frequen y domain belonging to model 12 and hara terizing well assigned time
behaviour are either slightly shifted to the left or frozen at their a tual lo ation.

Several eigenstru ture assignment equations orresponding to both models

being olle ted; it remains to nd the orresponding dynami gain (see 3.4.4,
Equation (3.31)) that will be in parallel with the initial one. The order of the
dynami gain is hosen as being as low as possible. During a trial and error
approa h it is found that a se ond-order ontroller would already provide good
performan e and de oupling, however, the matrix elements would be too high.
Therefore, a fourth-order ontroller is onsidered.

The state-spa e representation of the gain is given in [55. Bode plots of

ea h individual gain are also given in this report. Some gains appear to be high
pass. The resulting pole maps and the step responses an be found in Figures
19.3 (right) and 19.4 (bottom).

19.4.5 Inner lateral loop design

Let us now analyse the limitations that are related to the use of proportional
feedba k. In [55, two te hniques are used (multi-model as presented here
and pole migration as in [160). In both ases, our on lusions are similar:
proportional gains and damping ratio larger than 0.7 over all the ight domain
are in ontradi tion. The proposed proportional ontroller indu es a damping
ratio of about 0.4 (Figure 19.5, top-right) whilst time responses are perfe t
(Figure 19.6 top). When damping is augmented some poles be ome slower along
the real axis (Figure 19.5, bottom-left) and in turns, settling time be omes too
slow. A dynami feedba k, exa tly as proposed in 19.4.4 leads to Figure 19.5,
(bottom-right) for the poles and Figure 19.6 (bottom) for the time response.
These results are perfe t.

For the derivation of the proportional law the te hnique used is quite
straightforward; it onsists of assigning three slow poles (the same as for the
initial law, see 19.4.2) de oupled from  for the riti al model 12 (low speed,
high mass) and three fast poles (the same as for the initial law) de oupled from
 , based on the nominal model. Six eigenvalues are xed and six measurements
are available, therefore (3.22) an be solved without dynami s. The resulting

267
 Closed loop poles: lateral Closed loop poles: lateral with multimodel
4 4

3.5 3.5

3 3

2.5 2.5

2 2

1.5 1.5

1 1

0.5 0.5

0 0

0.5 0.5
4 3.5 3 2.5 2 1.5 1 0.5 0 0.5 4 3.5 3 2.5 2 1.5 1 0.5 0 0.5

Closed loop poles: lateral

Closed loop poles: lateral
4

4.5
3.5
4
3
3.5

2.5
3

2 2.5

2
1.5

1.5
1
1
0.5
0.5

0 0

0.5 0.5
4 3.5 3 2.5 2 1.5 1 0.5 0 0.5 5 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0

Figure 19.5: Lateral losed-loop poles. Top-left: initial design; top-right: pro-
posed proportional gain; bottom-left: well damped proportional gain but too
slow ; bottom-right: 4th order dynami gain

gain is:

 
2:45 1:96 2:32 3:52 0:27 1:51
Ki;lat;mult = (19.6)
3:96 0:11 4:81 0:45 1:34 0:14

19.5 Analysis of the Resulting Controllers

In this hapter further, mainly nonlinear veri ations, due to gust and ight
manuvres, will be presented. The simulations are made for the nominal
onguration.

19.5.1 Nonlinear simulation of an engine failure

Fig. 19.7: in ase of engine failure, the steady state deviation of the roll angle
is  = 4:3o = 0:075 rad, thus smaller than the expe ted 0:090 rad = 5o and is
redu ed to zero with an overshoot of 60 % after the restart of the failed engine,
ex eeding slightly the spe i ation of 50 %. Furthermore, the maximal value

268
 Beta versus demand of beta: 2^o Beta versus demand of phi: 20^o
2.5 1.5

2 1
Beta : deg

Beta : deg
1.5 0.5

1 0

0.5 0.5

0 1
0 20 40 60 0 20 40 60

Phi versus demand of beta: 2^o Phi versus demand of phi: 20^o
0.1 25

20
0
Phi : deg

Phi : deg
15

10
0.1
5

0.2 0
0 20 40 60 0 20 40 60
t:s t:s
Beta versus demand of Beta:2^o Beta versus demand of Phi:20^o
1
2
0.5
1.5

0
1

0.5 0.5

0 1
0 20 40 60 0 20 40 60

Phi versus demand of Beta:2^o Phi versus demand of Phi:20^o

0.3
20
0.2

0.1 15

0
10
0.1
5
0.2

0
0 20 40 60 0 20 40 60

Figure 19.6: Lateral step responses. Top: proposed proportional multi model
design (damping > 0.4 all over the ight domain), bottom: dynami multimodel
design (damping > 0.7 all over the ight domain)

269
ismax = 6:3o = 0:110 rad whilst up to 0:175 rad = 10o would be allowed (see
Fig. 19.7). As an be seen in the same gure, slideslip angle  is minimized.
The resulting lateral a eleration ny in reases up to 0:08 g. This passes the
normal level of 0:02 g but is still well within the limit of 0:2 g spe ied for
o
an engine failure. The heading rate has to remain less than 3 =s = 0.052 rad.
This stipulation is a tually met by _ max = 0:013 rad/s = 0:7o=s. Besides the
guarantee of performan e and omfort, the safety riteria are also respe ted.
The airspeed is

VA  77 m/s  62:2 m/s = 1:2 VStall (120 t)

and thereby, the angle of atta k meets with   1:7 = 0:03 rad the demand
o
of max  0:2 rad = 12 . Finally, the altitude is almost kept at 1000 m.
o

Measurements
1.5
Psi_dot [rad/s]
0.05
Psi [rad]

1.6 0

1.7 0.05
0 50 100 150 200 0 50 100 150 200
0.05
Beta [rad]

0.1
Phi [rad]

0 0
0.1
0.05
0 50 100 150 200 0 50 100 150 200
85 0.2
V_A [m/s]

ny []

80 0

75 0.2
0 50 100 150 200 0 50 100 150 200
0.2 1005
alpha [rad]

H [m]

0 1000

0.2 995
0 50 100 150 200 0 50 100 150 200
t [s] t [s]

Figure 19.7: The observation of the design riteria during engine failure

19.5.2 Nonlinear simulation observing safety riteria

Fig. 19.8 depi ts the results of the nonlinear simulations referring to safety
riteria on airspeed VA , angle of atta k  and roll angle . The manuvre
own orresponds to those during the landing approa h of the evaluation pro-
edure (see Chapter 19.6): engine failure, standard turn, des ent with a ight
path angle of 6o and 3o and wind shear. It is onsidered to be a good
representation of all arising ight onditions.

270
 Safety criteria
100

V_A [m/s]
90

80

70
0 50 100 150 200 250 300 350 400 450 500
0.2
alpha [rad]

0.2
0 50 100 150 200 250 300 350 400 450 500
1
Phi [rad]

0.5

0.5
0 50 100 150 200 250 300 350 400 450 500
t [s]

Figure 19.8: Safety riteria observation (the four segments are onsidered)

Generally speaking, the safety riteria are fullled:

Airspeed VA . The minimal airspeed is rea hed during the fourth segment with
about 75 m/s. Hen e, it is still 20 % larger than 1:2 VStall . For engine failure
see also se tion 19.5.2.

Angle of atta k. The maximal angle of atta k is dete ted during the turn but
 remains less than half of the limit of 0:2 rad = 12o. Results orresponding
the engine phase failure are given in Fig. 19.7.

Roll angle . The roll angle  remains well within all borders, ex ept during
the beginning of the standard turn. There, it ex eeds slightly, the limit of 0:5
rad = 30o. It is onsidered to be a eptable.

19.5.3 Nonlinear simulations under moderate turbulen e

onditions
In the following, the nonlinear simulations of the ontrolled air raft under mod-
erate turbulen e onditions are dis ussed. We have ug = vg = wg = 1:54
m/s = 0:1 W20 and Lug = Lvg = Lwg = 305 m, with altitude h = 1000 m
> 305 m and the wind speed at 20 ft above the ground W20 = 15:4 m/s for
moderate onditions.
With Fig. 19.9, a tuator eort minimisation is proved. Throttle rate _T H

271
 Actuator outputs and measurements
0.4 0.4

da_dot [rad/s]
0.2 0.2
da [rad]

0 0
0.2 0.2
0.4 0.4
0 50 100 0 50 100

dt_dot [rad/s]
0.2
0
dt [rad]

0
0.2

0.4 0.2
0 50 100 0 50 100
0.5 0.4
dr_dot [rad/s]

0.2
dr [rad]

0 0
0.2
0.5 0.4
0 50 100 0 50 100
dthr1_dot [rad/s]

0.15 0.02
dthr1 [rad]

0.1
0
0.05
0.02
0
0 50 100 0 50 100
dthr2_dot [rad/s]

0.15 0.02
dthr2 [rad]

0.1
0
0.05
0.02
0
0 50 100 0 50 100
t [s]
0.05
Phi [rad]

0
0.05

0 50 100
t [s]

Figure 19.9: A tuator dee tions and rates under moderate turbulen e ondi-
tions

272
remains well within 0:0042 rad/s = 0:15
0:028 rad/s = 0:15 _T H;max. The same
applies to aileron and rudder rates. They are normally less than 0:144 rad/s
= 0:33
0:436 rad/s = 0:33 _A;R;max. Only tailplane eort is slightly elevated.
Tailplane rate _T ex eeds - but not signi antly - the limit of 0:086 rad/s =
0:33
0:262 rad/s. However, a tuator deviations due to turbulen e are very
small. That means that there is still a lot of ontrol power left whi h ensures
manuveribility and agility. The maximum roll angle max in Fig. 19.9 is
about 1:5o = 0:025 rad whilst the spe i ation maximum is xed to 0:087 rad
o
= 5 .

Open loop Closed loop

0.04 0.04
RMS = 0.0120 RMS = 0.0097
0.02 0.02
beta [rad]

beta [rad]
0 0

0.02 0.02

0.04 0.04
0 50 100 0 50 100

0.03 0.03
RMS = 0.0094 RMS = 0.0144
0.02 0.02
Psi_c Psi [rad]

Psi_c Psi [rad]

0.01 0.01

0 0

0.01 0.01

0.02 0.02
0 50 100 0 50 100
t [s] t [s]

Figure 19.10: Comparison of  and responses, open and losed-loop, due to

unit RMS gust

In Fig. 19.10, additional disturban e redu tion hara teristi s of the pro-
posed ontroller are shown. Obviously, for the unit RMS lateral gust input the
losed-loop RMS sideslip angle RMS l = 0:0097
rad is smaller than the open-
loop (RMSol = 0:0120rad). On the other hand, the RMS heading angle error
RMS
= 0:0144 rad is higher than RMS
ol = 0:0094 rad, i.e. disturban es
l
are slightly redu ed for the inner loop quantities roll angle  and sideslip angle
 , but outer loop heading angle response is slightly deteriorated and gener-
ally speaking, natural behaviour is not hanged signi antly by the ontroller.
In ase of a failure in the ontrol system, this would be desirable. Nevertheless,
it would be better if the disturban e redu tion was in reased by the ontroller.

273
19.6 Results of the Automated Evaluation Pro-
edure
The landing s enario orresponding to four segments is des ribed in 14.3.3.

Segment I
See Figure 19.11 left side. Obviously, the ontroller fulls all these spe i a-
tions.

Segment II
See Figure 19.11 right side. The tra k ex eeds the bounds, sin e su essful
attempts to redu e the deviations have in reased the lateral a eleration ny
in an una eptable way and we are penalised for ex essive lateral a eleration
(explanation see se tion 19.2).

Segment III
See Figure 19.12 left side. The time responses remain within the maximal
bounds, unfortunately, they ex eed the limits in terms of agility. Basi ally, the
ontroller is too slow for the given bounds. But it was not possible to in rease
the gains without violating other onstraints.

Segment IV
See Figure 19.12 right side. Obviously the ee t of wind shear is very well
redu ed, almost evened out. It is possible to redu e the gains and to keep
the responses still within the bounds. This would also lead to smaller verti al
a eleration nz and hen e to better omfort. However, when the gains of the
perturbation redu tion loop are redu ed, the tra k following system be omes
inevitably slower whi h would result in even bigger deviations than we already
have (see Segments II/III).

Numeri al results
Table 19.2: All the numeri al values are a eptable ex ept for omfort indi a-
tors (see [55 for more details).

19.7 Con lusions

RCAM autopilot design is performed in two steps. The outer loop designs are
based on lassi al ontrol. For the inner loops, a new multi-model eigenstru -
ture assignment te hnique is proposed, it is briey des ribed in 3.4 and 3.4.4.
In the lateral ase the proposed multi-model ontroller is proportional, it is
omputed solving an equation similar to (3.29). However, the use of a dynami
gain is shown to lead to better results. In the longitudinal ase, a dynami

274
 First segment: top view Second segment: lateral deviations
300

100
200

50

lateral deviation [m]

100
xdeviation [m]

0 0
0 a b 1
1 c d 2

100
50

200
100

300
20 18 16 14 12 10 8 6 4 2 0 1 2 3 4 5 6 7 8
yposition (YE) [km] along track distance from point 1 [km]

Figure 19.11: Left: segment I - the ee t of engine failure. Right: segment II
- lateral deviations during the 3o/s turn

Third segment: altitude deviations Fourth segment: altitude deviations

30 30

20 20
altitude deviation [m]

3
altitude deviation [m]

10 2 f 3 10
4
g h
0 0

10 e 10

20 20

30 30
16 15 14 13 12 11 11 10 9 8 7 6 5 4 3 2 1
xposition (XE) [km] xposition (XE) [km]

Figure 19.12: Left: segment III - verti al deviations from the desired glideslope.
Right: segment IV - verti al deviations from the desired glideslope.

feedba k is proposed. It is omputed solving an equation similar to (3.31). In

pra ti e, gain s heduling is used for designing su h an autopilot. The proposed
dynami feedba k has a low order and does not need to be s heduled. The
degrees of freedom introdu ed by the dynami feedba k repla e with benets
those introdu ed by s heduling. In fa t, this is not a surprise, as the most
important parameter for robustness is the mass. It is also known that mass
variation indu es hanges in the dynami s, therefore the proposed autopilot an
be viewed as inherntly s heduled with respe t to mass variations via frequen y
variations.

Design y le. Any ontrol design te hnique is dened in order to deal, with
some priority, with a given subset of usual design riteria. For modal te h-
niques, the riteria that are naturally tra table are: settling time, overshoot,
damping ratio (see 19.3.1) and de oupling (see 19.3.2). Tra king (see 19.3.3)
was taken into a ount by using integrators. The hoi e of signals that are in-
tegrated was not analysed, we just reprodu ed the stru ture that is in use in
industry. Other riteria su h as omfort, fatigue... (see 19.3.5) were onsidered
mainly during the outer loop design by redu ing their dependen y on inner loop
design. More pre isely, adding several feedforward signals made possible the

275
 Segm. I Segm. II Segm. III Segm. IV Total

Perform. 0.3310 0.2516 0.7132 0.0464 0.3355

Perf. Dev. 0.0580 0.1585 0.0740 0.1230 0.1034
Comfort 0.4058 1.0015 1.3254 0.4756 0.8021
Safety 0.0080 0.0683 0.0062 0.0214 0.0260
Power 0.0030 0.0049 0.0120 0.0265 0.0116

Table 19.2: Numeri al results of the evaluation pro edure

use of slower inner loops; in turn, riteria of 19.3.5 ould almost be ignored.
Briey, the rst step of the design y le onsisted mainly of the hoi e of the
outer loop stru ture and of orrelated onstraints on erning the settling time
of the inner loops at the nominal point. Therefore, to deal with robustness,
the se ond step onsists of satisfying the above time domain onstraints but all
over the ight domain.

User friendliness of the method. The te hnique used here is illustrated in

3.4.4, Equ. (3.29)-(3.31). The required ontrol theory ba kground redu es
to the knowledge of the notion of modes (see 3.1). When dynami gains
are onsidered (longitudinal ase 19.4.4), numerous degrees of freedom are
introdu ed, most of them do not orrespond to eigenstru ture assignment on-
straints. Therefore they need to be optimized. The riteria we used are dened
in [150. In this ase, e ient numeri al tools are ne essary. Con erning the
multi-model feature of the proposed te hnique, a good ba kground in ight
me hani s (good sense also) is required. Indeed, when several obje tives are
dened for a given model, they are also more or less dened for adja ent mod-
els, hen e we must be very areful when additional obje tives on erning other
models are onsidered. It is very tempting to expe t oni ting results. To pre-
vent oni ting requirements we onsidered the most distant models (high mass
- low speed and low mass - high speed). Moreover, the hoi e of losed-loop
eigenve tors was based on orthogonal proje tions in order to ensure oheren y.
In the lateral ase, the proposed te hnique is mu h simpler be ause oni ting
obje tives are avoided (see 19.4.5).

E ien y for robustness. The proposed te hnique is very e ient, provided

that problems inherent in multi-model te hniques are well understood. After
having designed a good rst design, no more than one hour was required to treat
a new similar problem. To treat a problem for whi h no previous experien e is
available, a mu h longer analysis is ne essary to assess multimodel oni ting
obje tives.

Complexity of design. In the lateral ase, the ontroller redu es to a simple

proportional gain. In the longitudinal ase, the advantage of the proposed de-
sign te hnique is that dynami s are added only in order to treat robustness;
usually, dynami s are added to make solvable, a problem whi h would be oth-
erwise NP-hard. As a onsequen e, the number of states of the ontroller is

276
minimized (2 or 4 states).

Blo k diagrams. The used stru tures are given

 in Figure 19.2 with Ki in Equ. (19.6).

 in Figure 19.1 with Ki being a 4-dimensional dynami gain. Numeri al

values are given in [55.

Further work. Clearly, even if a proportional gain is proposed for the lateral
hannel, our on lusion is that the use of dynami feedba k is more e ient.
It permits us to repla e s heduled gains by simpler gains. However there are
two main drawba ks on erning the approa h used in this paper. The rst
problem is that it is di ult to tune our laws without oming ba k to the use
of design tools. The se ond problem is that we have not analysed the ee t
of our gains on the exible modes. But we believe that by using stru tured
gains (i.e. dynami s only between relevant inputs and outputs, with some
gains set to zero) both above problems are easily handled. Methods for solving
the stru tured gain multi-model eigenstru ture assignment problem are under
development and do not present any major theoreti al problem.

A knowledgement
The work presented in this hapter and the parti ipation of CERT-ONERA
in the denition of the RCAM design hallenge (14) was supported by the
Servi e Te hnique des Programmes Aronautiques (STPA).

277
20. The Lyapunov Approa h

Jamal Daafouz1;2 , Denis Arzelier1, Germain Gar ia1;2

and Ja ques Bernussou1

Abstra t. This hapter presents the dierent aspe ts of the design

methodology that has been used to ope with the RCAM design
problem. A quadrati approa h involving robust pole pla ement in
a disk and the solution of parameter-dependent Ri ati equations
is proposed. Robustness as well as performan e is a ounted for by
robust disk pole lo ation that ensures robust stability and provides
a good transient behaviour together with a spe i sele tion of the
weighting matri es involved in the Ri ati equations.

20.1 Introdu tion

Robust ontrol design has now rea hed a fairly high level of development and
industrial pa kages are available for su h purpose. Among the various methods
in robust ontrol, H1 optimization using state spa e models seems the most
popular. The reason being that, in its early stages, it appeared as a way to
extend the margin on epts used in the frequential single variable domain to
the multivariable one. Another method omes from the extension of Lyapunov
method to the study of un ertain systems, and it is termed quadrati approa h
in the literature.
This hapter is quadrati approa h oriented. This an be justied by the
fa t that the design obje tives are dened in the time domain. Moreover, when
dealing with multiobje tive problems, the H1 optimization approa h often on-
sists of dening a stru tured multi-blo ks un ertainty formulation for whi h the
stru tured eigenvalue tool is needed with the asso iated D-K iterations. Su h
algorithms still la k a strong onvergen e result so that their e ien y often
relies on the intuition of their users. The situation is a bit dierent for the
quadrati approa h whi h, at rst sight, presents a serious drawba k ompared
to the H1 one. The multiobje tive problems approa hed by single obje tive
with multiple onstraints problems are in fa t generally solved using only su-
ient onditions in the quadrati approa h. This is due to the fa t that a single
Lyapunov fun tion is involved in the onditions. However, this inferiority is,
in some sense, greatly redu ed by the fa t that the onditions possess, in most
ases, the onvexity property so that onstru tive numeri al methods an be

1
LAAS-CNRS, 7 avenue du olonel Ro he, 31077 Toulouse Cedex 4
2
Also with INSA, Complexe S ientique de Rangueil, 31077 Toulouse Cedex

278
proposed for their solution. For instan e, the LMI, (Linear Matrix Inequal-
ities), tools an be implemented when the onditions are expressed as ane
relations with respe t to these variables.

20.2 Sele tion of Controller Ar hite ture

From the Design Challenge spe i ations, the ontroller ar hite ture an be
hosen arbitrarily. It is however lassi al in air raft ontrol law design to arry
out a two-stage synthesis : one for the longitudinal motion and one for the
lateral motion leading to two separate ontrollers : a longitudinal ontroller
and a lateral ontroller. Ea h ontroller is designed separately.
Furthermore, ea h ontroller is then omposed of two ontrollers, (inner and
outer ontrollers), forming two loops : an inner loop and an outer loop. This is
also natural in air raft ontrol law design sin e the inner loop an be used for ro-
bust stabilization purposes while the outer is used for robust tra king purposes.
Figure 20.1 shows the ontroller ar hite ture used. K_lt_in and K_lg_in
are respe tively the lateral and longitudinal inner ontrollers. K_lt_out and
01
K_lg_out are the orresponding outer ontrollers. K l , K l , K g02 0 1 and K0 g2
are pre ompensators ne essary to have a unit stati gain.

20.2.1 Measurement signals

Longitudinal ontroller
As shown in gure 20.1, the longitudinal ontroller uses three measurement
signals in its inner loop : pit h rate q, z omponent of inertial velo ity in FV ,
wV and airspeed VA . This hoi e is onsistent with our hoi e of the ommands
in the inner longitudinal loop. In the outer longitudinal loop, the z position of
the air raft is used as measurement signal.

Lateral ontroller
The inner lateral loop uses two measurement signals : sideslip angle , roll
angle .

20.2.2 A tuator signals

Here, all the available a tuator signals are used, namely tailplane dee tion T ,
throttle positions of engine 1 and 2, T H1 ; T H2 , whi h are aggregated in one
signal, T H , are the longitudinal ontroller outputs. Aileron dee tion A and
rudder dee tion R , are the lateral ontroller outputs.

20.2.3 Filters and models

First of all, no lters of any kind are used or onsidered in our design pro e-
dure. Dierent kind of models have been used in our study : linearized models
and non linear model. The non linear model, the RCAM model, is dened

279
in the RCAM Design Challenge Manual, [145, and is a 6 degrees of freedom,
non-linear model of the air raft in landing onguration. This model is mainly
used in non-linear simulations in the analysis of the resulting ontroller in
terms of the applied methodology, (see se tion 20.5), and, of ourse, in the au-
tomated evaluation pro edure, (see se tion 20.6). The longitudinal and lateral
ontrollers are synthesized using un ertain linear models (see se tion 20.4).

Lateral Controller

K_lt_out

K_lt_in

.
c .
atan(80/g* c)

+ +
c +
+ A
yc
K 0l2 + K 0l1
+ c + Vv
+
+ R
Aircraft
+
Actuators
+
q
c T
zc + wc
+ q
K 0g 2 K 0g 1
+ Vc + wv
TH
+ V

K_lg_in

K_lg_out

Longitudinal Controller

Figure 20.1: Controller ar hite ture

280
20.2.4 Referen e signals
From the nine referen e signals available, the longitudinal ontroller uses:

2 3
z the altitude referen e
Lgref = 4 wV inertial verti al velo ity referen e 5
V total airspeed referen e
and the lateral ontroller uses:
 

y the lateral deviation referen e
Ltref = _
the rate of hange of tra k angle referen e

20.2.5 Resulting ontroller omplexity and implementa-

tion issues
Stru tural de oupling
As previously mentioned, a stru tural de oupling between longitudinal and lat-
eral design has been arried out. For ea h loop, a two-stage design from the
inner loop to the outer loop has been ondu ted.

Dynami order of the ontroller

The omputation of the dynami order of the ontroller takes the presented
design pro edure further into a ount. Indeed, for ea h motion, the design is
omposed of the following steps:

Linearized lateral model 5

Balan ed realization 4
A tuators 2
Integrators 2
Augmented plant 8 Inner ontroller 8
Extended plant 16
Balan ed realization 10 Outer ontroller 10
Global lateral ontroller 20
Table 20.1: Dynami order: lateral ontroller

- Extra tion of a linearized un ertain model.

- Constru tion of a linear augmented model, with in addition, the a tuators

and the integrators.

- Computation of an inner ontroller having the same dynami order as

the augmented plant.

281
 - Constru tion of an extended model omposed with the augmented one
and the inner ontroller.

- Computation of a balan ed realization of the previous extended model

for numeri al purposes.

- Computation of an outer ontroller having the same dynami order as

the redu ed extended plant.

This pro edure is the same for the longitudinal and the lateral motion ex ept
that it is ne essary to perform an additional redu tion for numeri al purposes
on the lateral linearized undertain model. This pro edure is detailed in se tion
20.4. Here is the summary of the dierent dynami orders involved in this
design.

Linearized longitudinal model 4

A tuators 2
Integrators 2
Augmented plant 8 Inner ontroller 8
Extended plant 16
Balan ed realization 10 Outer ontroller 10
Global longitudinal ontroller 20
Table 20.2: Dynami order: longitudinal ontroller

So, we get a nal global ontroller of dynami order 40 whi h has four
integrators. It is to be noti ed that no model redu tion method has been used
on the nal ontrollers whi h ould be easily performed and would provide a
great redu tion of this dynami order. In addition, it is lear from the design
pro edure that no s heduling has been used and that nonlinear ee ts are not
present in our ase.

Stability of the ontroller

As mentioned in the tutorial hapter, the methodology ensures that the re-
sulting ontroller will be stable. In fa t, the ontroller has the stru ture of an
observer. It an be shown, by using results from H1 theory that the ontroller
dynami matrix veries a Lyapunov equation whi h proves its stability.

282
20.3 Translation of Design Criteria into Method
Dependent Obje tives
20.3.1 Linear un ertain model
In this paragraph, we derive a linear un ertain model whi h takes into a ount
un ertainty due to mass and entre of gravity variations, that is:

m  M  m
 mass
x  x  x
y  y  y entre of gravity
z  z  z
The range of variations are des ribed in details in the RCAM manual [145.
Considering the extreme values of mass and entre of gravity, 24 extreme models
an be determined. Examining these models, an un ertain model is des ribed
where only un ertainty on matrix A has been onsidered. The global model
will be de omposed later into a lateral and a longitudinal un ertain model.
These models will be des ribed in the next se tion.

20.3.2 Steady state error

In order to satisfy the steady state error requirements, integrators will be added.
This is also presented with details in the next se tion.

20.3.3 Rise time and settling time

Inner loops
The requirements expressed in terms of time response, (rise time and settling
time), are dire tly taken into a ount by lo ating the poles appropriately in
the omplex plane. For the inner loops of the lateral and the longitudinal on-
trollers, the poles are lo ated in a disk entreed at 4 + j 0 with radius 3.6.
An absolute stability of 0:4 leads for the inner loops to a settling time about
ts ' 04::45 = 11s. On the other hand, the lo ation of some open loop poles (near
7 + j 0) is satisfa tory so the ontrol does not have to move them. This jus-
ties the hoi e of the disk entre at 4 + j 0 with radius 3:6.

Lateral outer loop

For the lateral outer loop, the ir le has to be hosen in a way whi h guaran-
tees a good tra king of referen e signal but also whi h preserves the dynami al
hara teristi s of the inner loop. After some trials, the disk entred at 10+ j 0
with radius 10 gave a good result. It is to be noted that, the disk orresponding
to the inner loop lies within the outer loop disk.

283
Longitudinal outer loop
The remarks for the longitudinal outer loop also apply here. A disk leading to
good results is entred at 6 + j0 with radius 5:84. The inner loop disk lies
also within this disk.

20.3.4 Other spe i ations

The other spe i ations are not handled expli itly. To take them into a ount
in the ontroller design, we use the degrees of freedom of the proposed approa h
for robust pole lo ation, that is the weighting matri es R1 ; R2 ; Q. A tuators
movement and engine power an be minimized by an appropriate sele tion of
the weighting matri es R1 ; R2 ; Q. In order to ontrol robustness margins
and wind ee ts, we have onstrained the sensitivity fun tions to be about 6
dB. This ensures a satisfa tory modulus margin .
1 We have onsidered the
sensitivity fun tions at the input of the a tuators and at the output of the
sensors.

20.4 Controller Derivation - Design Cy le

The design of the r am ontroller is de omposed into two separate subprob-
lems : a lateral hannel ontrol and a longitudinal hannel ontrol. For ea h
subproblem, an inner loop to ensure stabilization and an outer loop for tra king
are designed. The riterion onsidered in the design of the inner loop on erns
the magnitude of the sensitivity fun tions. The goal is to minimize perturba-
tion ee ts due to the wind. Limiting the sensitivity fun tion magnitude at
the input of the a tuators is a way to satisfy this requirement. In the outer
loop ase, the fundamental riterion is a good time response indu ing the de-
sired requirements on the settling time. This does not in lude overshoot. The
prin ipal steps in designing ea h hannel ontrol are :

i- Build the appropriate norm-bounded un ertain linear model. In lude

dynami a tuators and integrators and improve numeri al onditioning if
ne essary using balan ed realizations.

ii- For ea h loop, sele t a ir le in the left half plane and three weighting
matri es R1 ; R2 ; Q.
iii- Che k the magnitude of the sensitivity fun tions and time responses using
the linearized model.

iv- Che k robustness and performan e riteria on the losed loop non-linear
model.

If the result is unsatisfa tory, the design an be restarted from the step (ii-).

1
The modulus margin is dened as the radius of the ir le entred in [ 1; j 0 and tangent
to the Nyquist plot of the open-loop transfer fun tion.

284
20.4.1 Numeri al tools for ontroller synthesis
Matlab Simulink
Matlab
The synthesis and analysis are performed in a and envi-
ronment. To design the ontroller, a spe ial fun tion running under
environment is used to assign the system poles in a disk. This fun tion om-
putes a d-stabilizing ontroller in a ordan e with the algorithm proposed in

Matlab
se tion 5.3 of the tutorial hapter 5. The main toolbox used is Control tool-
box with . Some fun tions from the -Analysis Toolbox are used to
build losed-loop systems and in some ases to ompute balan ed realizations
of ill- onditioned systems.

20.4.2 Intermediate analysis

Lateral design model
In order to apply the d-stabilizing approa h, an un ertain linear model of the
form (20.1) is required.

x_ = (At + D1t F E 1t )x + Bt u
y = C t x + Dt u F 2F (20.1)

From the omplete linearized model we build the appropriate lateral design
model by pi king out the states, inputs and outputs appli able to the lateral
hannel :

2 3
p roll rate
6 r yaw rate 7
6 7
States : xlt = 6
6  roll angle 7
7
4 heading angle 5
vB y omponent of inertial velo ity
 
aileron defle tion
Inputs : ult = A
R rudder defle tion
2 3
 angle of sideslip
Outputs : ylt = 4  roll angle 5
vV y omponent of inertial velo ity

The resulting ABCD matri es are given by:

2
1:2667 0:5498 0 0 0:0242 3 2
0:8402 0:2904 3
0:0522 0:5207 0 0 0:0045 7 0 :0176 0:3325 7
Alt = 6
4 1 0:0284 0 0 05 Blt = 6
4 0 05
0 1:0004 0 0 0 0 0
2:2678 79:7679 9:7879 0 0:1699 0 2:0384

0 0 0 0 0:0125  
0 0
Clt = 0 0 1 0 0 Dlt = 0 0
0 0 2:2640 79:8667 1 0 0

285
From the un ertainty matri es of the omplete linearized model, we pi k out
the lines and olumns orresponding to the lateral omponents to get the lateral
hannel un ertainty domain:
2
0:2600 0 0:0420 3 
0 0 0:0103 7 1 0 0 0 0
D1lt = 6
4 0 0:1281 05 E 1lt = 0 1 0 0 0
0 0 0 0 0 0 0 1
7:4300 0 0
The design is started by using sideslip angle  and roll angle  as feedba k
signals. In order to improve numeri al onditioning we use a balan ed real-
ization of this un ertain model. A way to obtain this onsists of omputing
two balan ed realizations, one for the nominal model and the other one for the
un ertain model where the un ertainty is taken to be maximum ( F = 1). The
resulting un ertainty matri es ( D1 and E1 ) represent the dieren e between
these balan ed realizations. To obtain the lateral norm bounded un ertain
model, we add a tuator dynami s and integrators to the nominal balan ed re-
alization. We assume that there is no un ertainty on the a tuator dynami s or
on the integrators.

Longitudinal design model

Similarly to the lateral design, an un ertain linear model of the form (20.2) is
needed for applying the d-stabilizing approa h to the longitudinal hannel.

x_ = (Ag + D1g F E 1g )x + Bg u
y = Cg x + Dg u F 2F (20.2)

From the omplete linearized model, the appropriate longitudinal design model
is built by pi king out the states, inputs and outputs appli able to the longi-
tudinal hannel :
2 3
q pit h rate
6  pit h angle 7
States : xlg = 6 7
4 uB velo ity in body axis x dire tion 5
wB velo ity in body axis z dire tion
 
tailplane defle tion
Inputs : ulg = T
T H throttle position
2 3
q pit h rate
Outputs : ylg = wV z omponent of inertial velo ity 5
4
VA airspeed
The resulting ABCD matri es are given by:
2
0:9825 0 0:0006 0:0161 3 2
2:4379 0:2912 3
Alg = 4 1 0 0 0 Blg = 4 0 0
2:1937 9:7758 0:0325 0:0743 5 0:1837 9:8100 5
77:3570 0:7675 0:2264 0:6684 6:4785 0

286
 
1 0 0 0 
0 0
Clg = 0 79:8667 0:0283 0:9996 Dlg = 0 0
0 0 0:9996 0:0290 0 0
From the un ertainty matri es of the omplete linearized model we pi k out
the lines and olumns orresponding to the longitudinal omponents to get the
longitudinal hannel un ertainty domain:
2
0:0804 03 h
D1lg = 4 0 05 E 1lg = 0 0 0 1i
0 0 0 1 0 0
0 1:2851
The design is started by using the three outputs ( q; wv ; VA ) as feedba k
signals. As in the lateral design ase, we add the dynami a tuators and the
integrators to the previous model. Contrary to the lateral model, there is no
need to ompute balan ed realizations of the longitudinal model.

20.4.3 Design parameter adjustment strategy

A ir le in the left half-plane where the losed-loop poles should lie and three
weighting matri es ( R1 ; R2 ; Q) have to be hosen. The weighting fun tions
sele tion riterion is based on minimizing the magnitude of the sensitivity fun -
tions. The sensitivity fun tions onsidered here are the sensitivity fun tions at
the input of the a tuators and the sensitivity fun tions at the output of the
sensors. A way to a hieve this is to start the design by omputing a ontroller
with R1 = 1 1; R2 = 2 1; Q = 3 1 where 1 = 2 = 3 = 1. Then, iteratively
the parameters 1 ; 2 ; 3 are su essively adjusted and at ea h iteration a on-
troller is omputed. We stop when an a eptable magnitude of the sensitivity
fun tions is obtained. Here the adjustment of only one parameter namely 1
is su ient to a hieve sensitivity fun tions magnitudes of about 6 dB.

Lateral Channel - Inner loop

As explained in se tion 20.3, the lateral inner loop ontroller is su h that the
losed-loop poles lie in the disk with entre 4 + j 0 and radius r = 3:6. The
weighting matri es are : R1 = 1000  122 ; R2 = 122 ; Q = 188 . Taking
0 = 1, the solution is obtained for  = 3:8147:10 6 and the losed-loop poles
are shown in gure 20.2.
Figures 20.3-20.4 show the sensitivity fun tions and the omplementary sensi-
tivity fun tions of whi h the maximum is about 7:7 dB.

Lateral Channel - outer loop

The inner loop provides robust stabilization using ommanded sideslip  and

R  as inputs. For
roll angle lateral tra king we use the lateral tra king error

y = vV dt as the input. We start by omputing the inner losed-loop for
the nominal lateral model. An integrator is added to the third output vV in
order to re onstru t the lateral tra king error
y . The riterion in designing

287
 Poles location
4

Imag
0

4
8 7 6 5 4 3 2 1 0
Real

Figure 20.2: Closed-loop poles of the lateral inner loop

Sensitivity (Su) and compl. sensitivity (Tu) at the input Sensitivity (Sy) and compl. sensitivity (Ty) at the output
10
10

0
0
Singular values (dB)

10 Singular values (dB)

10
Su Sy
Tu Ty
20 20

30 30

40 40

50 2 1 0 1 2
50 2 1 0 1 2
10 10 10 10 10 10 10 10 10 10
Frequency (rad/s) Frequency (rad/s)

Figure 20.3: Sensitivity fun tions Figure 20.4: Sensitivity fun tions
of the lateral hannel at the in- of the lateral hannel at the out-
put of the a tuators ( Tmax = put of the sensors (Tmax =
4:91dB Smax = 5:78dB) 7:25dB Smax = 7:72dB)

the outer loop ontroller is to a hieve good time response : no overshoot and
settling time less than 45s. This is done by hoosing a ir le with entre -10+j0
R1 = 106  122 ; R2 =
and radius 10 and the weighting matri es are xed as :
1; Q = 11010 . Taking 0 = 1, the pro edure su eeds with  = 3:7253 10 9 .

Longitudinal Channel - Inner loop

The longitudinal inner loop ontroller is su h that the losed-loop poles lie in
the disque with entre 4 + j 0 and radius r = 3:6. The weighting matri es
are :R1 = 104  122 ; R2 = 133 ; Q = 188 . Taking 0 = 1, the solution is

obtained for  = 2:44 10
4 and the losed-loop poles are shown in gure 20.5.
Figures 20.6-20.7 show the sensitivity fun tions and the omplementary sensi-
tivity fun tions of whi h the maximum is about 7 dB.

Longitudinal Channel - outer loop

To a hieve altitude tra king, the design of an outer loop ontroller is onsidered.
An integrator is added to the inner losed-loop to obtain the altitude output z

288
 Closed loop poles
4

Imag
0

4
8 7 6 5 4 3 2 1 0
Real

Figure 20.5: Closed-loop poles of the longitudinal inner loop

Sensitivity (Su) and compl. sensitivity (Tu) at the input Sensitivity (Sy) and compl. sensitivity (Ty) at the output

10

0
0

10
10
Singular values (dB)

Singular values (dB)

Su
Sy
Tu
Ty
20
20

30 30

40 40

50 2 1 0 1 2
50 2 1 0 1 2
10 10 10 10 10 10 10 10 10 10
Frequency (rad/s) Frequency (rad/s)

Figure 20.6: Sensitivity fun tions Figure 20.7: Sensitivity fun tions
of the longitudinal hannel at the of the longitudinal hannel at the
input of the a tuators ( Tmax = output of the sensors ( Tmax =
4:64dB Smax = 4:89dB) 4:21dB Smax = 7:11dB)

from the z- omponent of the inertial velo ity wv . The design of the outer loop
is now a single input-single output problem sin e we only use the ight path
ommand to a hieve the altitude tra king. The riteria in designing the outer
loop ontroller is to ensure good tra king result in terms of time response : no
overshoot and settling time less than 45 s. A ir le entreed at 6 + j 0 with
radius r = 5:84 and the weighting matri es set to: R1 = 40; R2 = 1; Q =
11010 give an a eptable losed-loop altitude
5.
response. Taking 0 = 1, the
solution is obtained for  = 6:1035 10

The ontroller design is now omplete. The analysis of the resulting ontroller
in terms of the applied methodology is the purpose of the next se tion.

20.5 Analysis of the Resulting Controller

The analysis of the resulting ontroller is twofold. First, we have performed
an analysis on the nominal linearized models presented in the se tion 20.4,
separating longitudinal and lateral motions. Se ondly, a non linear analysis
has been performed using the design environment des ribed in [145.

289
 20.5.1 Linear analysis
The performan e of the ontroller is he ked by analysing the ommand re-
sponse hara teristi s to step referen e signals. The ommand response har-
a teristi s are dened in terms of rise time tr , settling time, ts and overshoot
Mp . These hara teristi s are dened in [145.

Lateral deviation
The response to a step of 1 m in the lateral position is given in gure 20.8.
The requirements for rise time, ( tr = 11:33 s), settling time, ( ts = 28:56 s),
and overshoot, ( Mp = 0 %) are met.

Lateral response Step response to an altitude command

1
1

0.8
0.8
Laterl position (m)

Altitude (m)

0.6

0.6

0.4

0.4
0.2

0.2
0

0 0.2
0 5 10 15 20 25 30 35 40 45 0 5 10 15 20 25 30
Time (s) Time (s)

Figure 20.8: Closed loop lateral re- Figure 20.9: Altitude response to a
sponse to a unit lateral ommand step in altitude of 1m

Altitude response
The tra king of altitude ommands has been tested by plotting the altitude
response to a step in altitude of 1 m (gure 20.9). The requirements for rise
time, ( tr = 6:92 s), settling time, ( ts = 16:07 s), and overshoot, ( Mp = 0 %)
are met.

Flight path angle response

The system is subje ted to a step of 3 deg in ommanded ght path angle (gure
20.10). The requirements on rise time, ( tr = 5:26 s), settling time, (ts = 11:87
s), and overshoot, ( Mp = 0 %) are met. The ross oupling between ight path
angle and velo ity is weak (maximum value of the absolute error of V of 0.04
m/s). The elevator and engine ommands are also given and appear to be not
too large. Similarly, a response to a step of 13 m/s in ommanded airspeed
leads to : tr = 6:65 s, ts = 13:72 s and Mp = 0 %.

290
 4 Response to roll angle command
25

path angle (deg)

phi
3
20

2
15

Angle (deg)
10

V (m/s)
0
5
de (deg)
beta
1 0

thr (deg)
2 5 dr

da
3 10
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
Time (s) Time (s)

Figure 20.10: Response to a step in Figure 20.11: Response to a 20 deg

ight path angle of 3 deg step in roll angle ommand

Roll angle response

The gure 20.11 presents the response of the system in terms of roll angle and
sideslip angle to a 20 deg step in roll angle ommand. The ross oupling from
roll angle to sideslip angle is less than 0:17 deg. The hara teristi s of the
response are tr = 4:18 s, ts = 9:26 s and Mp = 0 %. Jointly to the response
to a step in roll angle, the response to a step of 2 deg in sideslip ommand is
hara terized by : tr = 4:55 s, ts = 9:21 s and Mp = 0 %.

20.5.2 Robustness riteria

Centre of gravity variations and mass variations
Robustness spe i ations are dire tly taken into a ount by the norm-bounded
un ertainty modelling, (see the un ertain lateral and longitudinal models). If
the algorithm su eeds, the dedu ed ontroller, using the quadrati d stabi-
lizability on ept, ensures robust stability and performan e against entre of
gravity variations and mass variations in the range des ribed by the un ertain
modelling. This range overs those des ribed by the robustness riteria. Here,
minimum performan e is a hieved by the robust pole pla ement in a disk. So,
this property is he ked by s anning the losed-loop poles of the linearized
models when the un ertainty term F varies in F (gures 20.2 and 20.5).
Velo ity variations
Robustness is also assessed by onsidering a set of linearized models dened in
table 20.3. This set ontains the nominal model M0 and other models obtained
by the trimr am pro edure for dierent values of mass, entre of gravity and
velo ity. Models M1 , M2 and M3 are the ones spe ied in the addendum to
the RCAM Design Challenge [143.
Responses of the system in terms of sideslip angle for the lateral hannel and
velo ity for the longitudinal hannel are shown in gures 20.12 and 20.13. For

291
 Model M0 M1 M2 M3 M4 M5 M6 M7 M8 M9
Mass 120 100 125 150 100 100 100 100 100 110
Cxg 0.23 0.31 0.31 0.31 0.31 0.31 0.15 0.31 0.15 0.31
Czg 0 0 0 0 0.21 0 0 0 0.21 0
Va (m/s) 80 58 90 71 58 71 80 90 90 58
Model M10 M11 M12 M13 M14 M15 M16 M17 M18 M19
Mass 110 120 120 120 130 130 140 140 150 150
Cxg 0.31 0.31 0.31 0.31 0.31 0.31 0.31 0.31 0.31 0.15
Czg 0 0 0 0.21 0 0 0 0 0.21 0.21
Va (m/s) 80 58 71 90 71 90 71 90 71 90
Table 20.3: A set of models to he k robustness

1 1
V (m/s)
Beta
0.8 0.8

0.6 0.6
Angle (deg)

0.4 0.4

0.2 0.2

0 0

Phi Path angle (deg)

0.2 0.2

0.4 0.4
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40
Time (sec) Time (sec)

Figure 20.12: Response to a 1 deg Figure 20.13: Response to a step in

step in sideslip angle Velo ity of 1 m/s

the onsidered set of linearized models, the system remains stable and perfor-
man es are well preserved.

Time delay
To he k robustness with respe t to time delay variations, all the previous
simulations have been performed with a delay equal to 50 ms and with a time
delay equal to 100 ms. As an example, the augmentation of the delay does not
ae t the hara teristi s of the response of the sideslip angle, the roll angle or
the ross oupling between these two variables.

20.5.3 Non linear analysis

The non linear analysis has been ondu ted in a similar way to the linear one.
Here, of ourse, the simulation environment we have onsidered is the one given
in [145. The general trim onditions for the non linear simulations are given
in the table below :
It is to be noted that onstant wind speed and turbulen e are not onsidered

292
 Mass 120000 kg
Centre of gravity x-pos. 0.23
Centre of gravity y-pos. 0
Airspeed 80 m/s
Inertial ight path angle 0 deg.
Inertial tra k angle -90 deg.
Initial position in FE [0 0 -1000 m
Computational time delay 0 s
S aling st. deviation gust 15.4 m/s
Constant wind speed [0 0 0 m/s

Table 20.4: Trim onditions for the non linear analysis

here ex ept where expli itly mentioned.

Performan e riteria
The simulation parameters and the results of the time simulations, rise time,
settling time and overshoot are given here. The later hara teristi s are indi-
ated on the plots by dashed lines.

Altitude response
The tra king of altitude ommands has been tested by plotting the altitude
response to a step in altitude of 50 m (gure 20.14). The hara teristi s of the
response are tr = 4:32 s, ts = 15:9 s and Mp = 2:13 %.

940 76

950 78

960 80
Heading angle (deg)
Altitude (m)

970 82

980 84

990 86

1000 88

1010 90
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Time (s) Time (s)

Figure 20.14: Altitude response to Figure 20.15: Heading angle re-

a step in ommanded altitude of sponse to a step in ommanded
50 m heading angle of 13 deg

293
 Heading angle response
The tra king of heading angle is illustrated in the gure 20.15 where the re-
sponse to a step in ommanded heading angle of 13 deg is represented. The
hara teristi s of the response are tr = 4:4 s, ts = 10:7 s and Mp = 0 %.

Roll angle response

An engine failure is simulated at t = 10 s and the failed engine is restarted at
t = 100 s (gures 20.16 and 20.17) .The maximum value of the absolute error
of the roll angle  is of 3:82 deg. There is no steady state error due to the
engine failure. After the restart of the failed engine, the maximum value of the
absolute error of the roll angle  is of 5:56 deg. After the engine failure, the
maximum value of the absolute error of the sideslip angle  is of 2:7 deg. After
the restart of the failed engine, the maximum value of the absolute error of the
sideslip angle  is of 4:27 deg.
6
10

8
4
6

4
2
Sideslip angle (deg)
Roll angle (deg)

0 0

4 2

6
4
8

10
6
0 20 40 60 80 100 120 140 160 180 200 0 20 40 60 80 100 120 140 160 180 200
Time (s) Time (s)

Figure 20.16: Roll angle response Figure 20.17: Sideslip angle re-
sponse

Both throttle responses are shown here (gure 20.18). In gure 20.19, roll angle
response is represented for moderate turbulen e, that is, we have sele ted wind
at 20 ft above the ground, W20 = 15:4 m/s.

Heading rate
Here, engine failure is observed at t = 10 s and engine is starting again at
t = 50 s (gure 20.20). The peak maximum heading rate, _ is equal to 0:0349
deg/s.

Cross oupling between airspeed VA and altitude h:

The ross oupling between h and VA is illustrated in the gure 20.21. For a
step in velo ity ommand of 13 m/s, we get a maximum value for the absolute

294
 0.18 6

0.16 Throttle 2

4
0.14

0.12 2
Throttle 1 2 (rad)

Roll angle (deg)

0.1
0
0.08

0.06 2

0.04
4
0.02 Throttle 1

0 6
0 20 40 60 80 100 120 140 160 180 200 0 10 20 30 40 50 60 70 80 90 100
Time (s) Time (s)

Figure 20.18: Throttle responses Figure 20.19: Roll angle response

for moderate turbulen e

0.04 986

988
0.03

990
0.02

992
Psi_dot (deg/s)

Altitude (m)

0.01

994

0
996

0.01
998

0.02
1000

0.03 1002
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Time (s) Time (s)

Figure 20.20: Heading rate re- Figure 20.21: Altitude response for
sponse for an engine failure a step in airspeed of 13 m/s

error of the altitude h equal to 12:93 m. For a step in altitude of 30 m, we get

a maximum value for the absolute error of V of 0:23 m/s.

20.6 Results of the Automated Evaluation Pro-

edure
This se tion presents the methodology independent results of the ontroller
designed in the previous se tions. It is mostly based on the evaluation mission
and s enario dened in [145 : both overall tra king performan e and inner-
loop behaviour of the ontrolled system will be evaluated by means of bounds
on key variables.

The ee t of engine failure

As the RCAM air raft model is twin-engined, a single engine failure will mainly
result in lateral deviation. Hen e gure 20.22 provides a top view of the rst
traje tory segment. The stati lateral deviations satisfy all the required on-
straints with very small overshoots orresponding to engine failure and engine

295
 restart.

First segment: top view Second segment: lateral deviations

300

100
200

50

lateral deviation [m]

100
xdeviation [m]

0 0
0 a b 1
1 c d 2

100
50

200
100

300
20 18 16 14 12 10 8 6 4 2 0 1 2 3 4 5 6 7 8
yposition (YE) [km] along track distance from point 1 [km]

Figure 20.22: The ee t of engine Figure 20.23: Lateral deviations

failure during the 3 deg/s turn

The 3 deg/s turn

At the beginning of the turn, the perfe t following of the required traje tory
and the desire to perform a oordinated turn would imply a sudden hange
in the air raft's bank angle, whi h is only possible with an innitely high roll
rate. Obviously this is undesirable but deviations from the desired traje tory
at the start, (and the nish), of the turn are unavoidable. As it an be seen
on the gure 20.23, whi h is a loser look at the a tual lateral deviations, the
traje tories are onned within the xed bounds.

The apture of the -6 and -3 degrees glideslope

We start with a glideslope of -6 deg; again it is unavoidable that the air raft
leaves the desired traje tory. It returns to the traje tory without overshoot
and well within a period of 30 s. After that, we go to a glideslope of -3 deg
su h that we get an inverse behaviour with respe t to the desired traje tory,
that is about half the size of the rst response. The verti al deviations from
the desired glideslope are plotted in gure 20.24.

The nal approa h with windshear

While on nal approa h with a glideslope of -3 deg the ee t of a windshear
model is onsidered. The verti al deviations from the desired glideslope are
plotted in gure 20.25. The behaviour of the ontrolled system is satisfa tory.

Numeri al results
Table 20.5 summarizes the numeri al results obtained with the designed on-
troller in terms of performan e indi ators. For the motivation and al ulation
prin iple of the various results see [145.

296
 Third segment: altitude deviations Fourth segment: altitude deviations
30 30

20 20

altitude deviation [m]

altitude deviation [m]

3
10 2 f 3 10
4
g h
0 0

10 e 10

20 20

30 30
16 15 14 13 12 11 11 10 9 8 7 6 5 4 3 2 1
xposition (XE) [km] xposition (XE) [km]

Figure 20.24: Verti al deviations Figure 20.25: Verti al deviations

from the desired glideslope from the desired glideslope

Segm. I Segm. II Segmm. III Segm. IV Total

Performan e 0.1753 0.1058 0.5940 0.1698 0.2612
Perf. Dev. 0.0888 0.3484 0.1603 0.4737 0.2678
Comfort 0.5011 16.7323 1.3541 0.6029 4.7976
Safety 0.0056 0.1544 0.0054 0.0365 0.0505
Power 0.0055 0.0252 0.0139 0.0278 0.0181

Table 20.5: Numeri al results of the evaluation pro edure

All the indi ators ex ept omfort are less than 1 (maximal bound spe ied
in [145) indi ating a relatively good behaviour of our ontroller. In fa t, we get
non a eptable value of the lateral a eleration be ause in segment II, (whi h
orresponds to a turn of 90), the ontrol is not smooth enough. Lo ating the
inner loop poles in a ir le larger than the previous one provides a smooth
ontrol in the lateral hannel. A redesign of the lateral ontroller has been
onsidered. A lateral inner ontroller has been omputed to lo ate the poles in
a ir le entered at 8 + j 0 with radius r = 7:9 and with the same weightings
matri es as previously. It is not ne essary to hange design parameters of the
outer loop ontroller. An outer loop ontroller is obtained with the same hoi e
of ir le and weightings matri es as in the previous design. The evaluation
pro edure has been run with the new ontroller. Lateral deviations are shown
in gures 20.26 and 20.27 and the new numeri al results are given in table
20.6. The omfort indi ator has been signi antly improved by hanging only
one design parameter, namely the ir le. We believe that a renement in the
lateral inner ontroller design to get omfort indi ator less than 1 is possible
by hanging other design parameters, namely weightings matri es.

20.7 Con lusions and Lessons Learned

After a phase of analysis and evaluation, the results produ ed by the use of the
designed ontroller appear to be quite good. The main strong features of the

297
 First segment: top view Second segment: lateral deviations
300

100
200

50

lateral deviation [m]

100
xdeviation [m]

0 0
0 a b 1
1 c d 2

100
50

200
100

300
20 18 16 14 12 10 8 6 4 2 0 1 2 3 4 5 6 7 8
yposition (YE) [km] along track distance from point 1 [km]

Figure 20.26: The ee t of engine Figure 20.27: Lateral deviations

failure during the 3 deg/s turn

Segm. I Segm. II Segm. III Segm. IV Total

Performan e 0.2053 0.1380 0.2115 0.2864 0.2103
Perf. Dev. 0.0748 0.0937 0.1500 0.2484 0.1417
Comfort 0.5185 3.8774 1.3632 0.4857 1.5612
Safety 0.0068 0.2426 0.0088 0.0586 0.0792
Power 0.0046 0.0080 0.0150 0.0309 0.0146

Table 20.6: Numeri al results of the evaluation pro edure

methodology an be established in the following way.

- It does not require any pre ise knowledge or prior experien e about ight
ontrol systems design. This point an be illustrated by the fa t that it
is the rst appli ation of this methodology in the aeronauti al eld by
the team of the LAAS-CNRS.

- This methodology is well suited to the two-stage design, whi h appears

to be inherent to the ight ontrol systems design.

- The use of the methodology does not need a very large number of synthe-
sis parameters whi h are learly and easily identiable. A tions on these
parameters an be translated in lear frequential interpretations.

- The pro edure of synthesis an be easily automated.

- The numeri al tools whi h have been used are reliable and widespread in
the s ienti world.

- The stru ture of the ontroller is simple and natural for su h problems.

- The time spent on designing the ontroller was not too important. Most
of the spent time was used to understand the problem and to dene the
ar hite ture of the ontroller.

On the other hand, some drawba ks need some improvements:

298
 - The dimension of the resulting ontroller may be large when applying a
multistage design whi h is the ase here with the inner and outer loops.
Indeed, for ea h design a ontroller is obtained with a dimension equal
to the one of the model used. Getting a redu ed order ontroller may be
a hieved by using standard expli it model redu tion method but with the
known di ulties linked with the ne essity of ontrolling the performan e
degradation.

- It is ne essary to have a pre ise model of the un ertainty in order to

as ertain the robust stability and pole lo ation.

- If un ertainty had to be onsidered in other terms of the model, (B; C; D),

the parameter-dependent Ri ati equations ould not be still available.
The Ri ati type approa h in quadrati design is orrelated to spe i
un ertainty formulation su h as the norm-bounded one and does no longer
hold when interval-type or polytopi un ertainty is onsidered. An LMI
formulation should be onsidered.

Possible extensions and improvements are possible, mainly the size of the on-
troller an be redu ed by using expli it model redu tion methods. Moreover,
other ar hite tures, (use of other measurement signals...), for the ontroller
should be explored.

299
21. An H1 Approa h
Mark R. Tu ker and Daniel J. Walker
1 2

Abstra t. This hapter des ribes an H1 approa h to the design

of a ontroller for the Resear h Civil Air raft Model (RCAM). The
ontroller produ ed onsists of an inner and outer loop. The in-
ner loop ontrols the manual ying of the plane and is designed
using a multivariable H1 two degree-of-freedom mixed sensitivity
approa h. The design in orporates performan e and robustness re-
quirements to produ e a losed loop de oupled system to meet the
ying spe i ations. The outer loop deals with autopilot tra king
and is designed using H1 normalised oprime fa tor one degree
of freedom loop shaping te hniques. Three outer loop ontrollers
are synthesised to tra k height, heading and to reje t lateral devi-
ations. Analysis tests and simulation results look promising. The
ontrollers produ ed are of a high order, although in prin ipal it
should be possible to a hieve signi ant order redu tion. For a
predened mission s enario, the performan e, robustness, omfort,
safety and power results satisfy the spe i ation.

21.1 Introdu tion

In this hapter, the design methodology used for the RCAM ben hmark prob-
lem is based on H1 ontrol theory. Sin e the 1980's, H1 ontrol theory has
been establishing itself as as powerful te hnique for the design of multivariable
ontrollers that are robustly stabilising as well as meeting performan e require-
ments. It is a frequen y domain te hnique whi h an a ommodate robustness
issues in luding disturban e reje tion and model un ertainty. Frequen y de-
pendent weights are sele ted to a ommodate robustness and performan e re-
quirements. The losed loop system an also be mat hed to an ideal system by
in luding mat hing models into the design.
Owing to variations in speed, weight, entre of gravity, time delays, engine
failure and disturban es from wind shear and gusts, and the need to meet the
performan e spe i ations dened in [145, the RCAM problem is very suitable
to the appli ation of H1 te hniques.
1
Engineering Department, University of Lei ester, University Road, Lei ester LE1 7RH,
United Kingdom. E-mail: mrtsun.engg.le.a .uk Tel: +44 116 252 2567/2874 Fax: +44 116
252 2619
2
Engineering Department, University of Lei ester, University Road, Lei ester LE1 7RH,
United Kingdom. E-mail: wjdlei ester.a .uk Tel: +44 116 252 2529 Fax: +44 116 252 2619

300
 The subsequent se tions of this hapter detail the hosen ar hite ture and
how the design requirements are in orporated into the design y le. Two dis-
tin t H1 te hniques are used. Analysis of the resulting ontrollers and evalu-
ation results are then given.

21.2 Controller Ar hite ture

The ar hite ture used is based on an inner and an outer loop stru ture (Figure
21.1). The two loops are employed to separate the manual ying requirements,
with the pilot in the loop, from the tra king of the desired traje tory when in
autopilot mode. A number of sele tors exist, so that the ontroller an swit h
between dierent modes; eg full authority autopilot to manual.

Outer Loop
vx_d _d
Demanded Inner Loop
vy_d
Track Angle
_d
Outer Heading
Controller

ylat Lateral Deviation

Controller V_d
Inner Loop Gain Scheduling RCAM Model
Controller For Speed and Actuators
z_d

Outer Height q, vz, Vair, beta, p, r, phi

wv _d
Controller
z

wv _d

ylat ,, z

Figure 21.1: RCAM Controller

21.2.1 Inner Loop

The inner loop is simply one multivariable ontroller designed to meet the
manual ying performan e and robustness spe i ations. Rather than design-
ing separate longitudinal and lateral ontrollers one ombined ontroller has
been designed. The H1 optimisation allows a single ontroller to be designed
expli itly so that the ross- oupling error between the ontrolled hannels is
redu ed.

Referen e Signals
The referen e signals to the inner loop are the verti al velo ity demand ( wV _d),
airspeed demand ( VA_d ) and roll angle demand ( _d ).  These signals are
adequate for manual ying and an be used by the outer loop to tra k the
height and heading as well as reje ting lateral deviations.

301
Measurements Signals
The measurements signals used by the inner loop are the measured verti al
velo ity (wV ), airspeed (VA ) 
and roll angle ( ), as well as the pit h rate ( ), q
r p 
yaw rate ( ) and roll rate ( ) and sideslip ( ). The three rates are fed ba k to
enhan e the ontrol and stabilisation of the inner loop. The sideslip ( ) is fed 
ba k to be ontrolled to zero.

A tuator Demands
The aileron ( A ), tailplane ( T ), rudder ( R ) T H )
and engines ( are all on-
trolled by the inner loop, with the two engines ontrolled as one. For robustness
to airspeed variations, the ontrol surfa e a tuator signals are gain s heduled
to a ount for the fa t that the ontrol surfa e ee tiveness is a fun tion of the
square of the airspeed.

21.2.2 Outer Loop

The outer loop omprises three separate ontrollers, used spe i ally for tra k-
ing a traje tory when in autopilot mode.

Height Tra king

The height tra king outer loop ontroller has three inputs; height demand,
height rate demand and the feedba k of the a tual height. The ontroller
generates a verti al speed demand for the inner loop. Note that the referen e
height rate demand input is swit hed in for tra king ramped inputs, and is
swit hed out if stepped demands are required.

Inertial Tra k Angle Tra king

The inertial tra k angle ontroller has three inputs; inertial tra k angle demand,
inertial tra k angle rate demand and the feedba k of the a tual inertial tra k
angle. The ontroller generates a roll angle demand for the inner loop to
o-ordinate a turn. Note that the referen e inertial tra k angle rate demand
input is swit hed in for tra king ramped inputs, and is swit hed out if stepped
demands are required.

Lateral Deviation Reje tion

This ontroller takes the lateral deviation error as the input signal and orre ts
for this deviation by generating an inertial tra k angle demand. The inertial
tra k angle demand is dire tly fed into the outer loop heading referen e om-
mand whi h generates a roll angle demand to oordinate a turn to orre t for
the deviation.

302
21.3 Translation of RCAM Design Criteria
The two degree of freedom mixed sensitivity design for the inner loop allows for
the in lusion of some design riteria dire tly into the design methodology. A
number of weights are used in the design and many of these are derived dire tly
from the design riteria. The outer loop design is mu h simpler in omparison,
and design goals are simply met by tuning the shaping parameters.

21.3.1 Performan e Spe i ation

From the spe i ation, a mat hing model for the referen e signals an be on-
stru ted for in lusion in the design.

Verti al Velo ity

The verti al velo ity ( wV ),
inertial ight path angle ( ) and inertial velo ity
V ) are related by
(

wV
sin( ) = (21.1)
V

As the ight path angle ( ) is typi ally small, sin( )  , so for onstant
V)
inertial velo ity (
wV / (21.2)

Thus the ight path angle spe i ation an be used as the verti al velo ity
spe i ation. To meet the rise time ( tr ), overshoot (ts ) and settling time (Mp )
spe i ations, the following se ond order model is hosen

!n2 1
M1 = 2
s + 21 !n1 s + !n2 1
(21.3)

with !n1 = 0:85 rad/s and 1 = 1. This model a hieves the following results,
with the spe i ation requirements shown in bra kets for omparison.

tr = 4:0s (< 5s) ts = 7:8s (< 20s) Mp = 0:0% (< 5%) (21.4)

Airspeed
The rise time, overshoot and settling time spe i ation for airspeed is also
given. Using the mat hing model

!n2 2
M2 = 2
s + 22 !n2 s + !n2 2
(21.5)

with !n2 = 0:3 rad/s and 2 = 1 yields the following results, with the spe i-
ation requirements on e again shown in bra kets for omparison.

tr = 11:2s (< 12s) ts = 22:1s (< 45s) Mp = 0:0% (< 5%) (21.6)

303
Roll Angle

For a oordinated turn, the roll angle ( ) an be approximated in terms of the
V ) and heading rate (_ ) by
inertial velo ity (
 
V _
 = atan (21.7)
g
Hen e for small angles
 / _ (21.8)

The spe i ation gives the ideal heading response, so to ontrol the roll angle
(whi h is proportional to heading rate), the heading rate bandwidth should be
2-5 higher than for the heading. In terms of the time domain, the heading
response rise time is tr < 10 se s, so the heading rate rise time should be 2 5
se s. The ideal se ond order model hosen was

!n2 3
M3 = 2
s + 23 !n3 s + !n2 3
(21.9)

with !n3 = 1 rad/s and 3 = 1, leading to

tr = 3:4s ts = 6:6s Mp = 0:0% (21.10)

Cross-Coupling
Using the se ond order systems 21.3, 21.5 and 21.9 the ideal system M is formed
as the transfer fun tion matrix

M = diag s2 +10:7:7225 0:09 1
s+0:7225 ; s2 +0:6s+0:09 ; s2 +2s+1 (21.11)

All o-diagonal elements are zero, orresponding to zero ross- oupling between
the hannels.

21.3.2 Robustness
Robustness to the plant un ertainties of entre of gravity and mass variations
are not expli itly in orporated into the design. It is noted that su h parameter
variations ould be modelled as additive perturbations and in orporated into
the design weights. Here however, the parameter variations are only onsidered
for sele ting a nominal ight ondition for the RCAM linearisation. The linear
model was produ ed for air raft mass of 120000kg, entre of gravity in x of
0:23m and entre of gravity in z of 0:0m.
The following robustness issues are expli itly in luded in the design.

Time Delays
For the time delays, a nominal delay of 0:05s was hosen for ea h of the ontrol
signals and was represented in the model as rst order Pad approximations:
i.e.
40 s
(21.12)
40 + s

304
Wind and Output Perturbations
The ontroller is designed to be robust to output perturbations and wind dis-
turban es. The design takes advantage of the RCAM model's wind inputs.
The inputs are used in the ontroller synthesis to minimise the ee t of wind
disturban e.

Airspeed
As ontrol surfa e ee tiveness is a fun tion of the airspeed squared, the RCAM
is made robust to airspeed variations by gain s heduling the ontrol surfa e
demands to aileron ( A ), T )
tailplane ( and rudder ( R ) as a fun tion of the
airspeed squared.

21.4 Design Cy le
The design y le onsists of rst the inner, then the outer loop design. The inner
loop design is based on a linearization taken from the non-linear RCAM model,
whi h is inserted into a spe ially sele ted system design stru ture. Weights are
then hosen. Having synthesised an H1 sub-optimal ontroller, ontroller re-
du tion, then testing are performed. The outer loop design pro eeds in a simi-
lar manner, ex ept the linear models used are derived from the ideal mat hing
models used for the inner loop design.
Further testing is performed on the omplete system, and iterations of the
design y le are performed.
This next se tion des ribes the inner and outer loop H1 designs3 in more
detail.

21.4.1 Inner Loop Design

Linearization
The model used for the ontroller synthesis omprises three subsystems: the
linearization of the RCAM rigid -body dynami s, an a tuator linearization and
a time delay Pad approximation. The RCAM itself is linearised around the
operating onditions of mass ( 120000 kg), entre of gravity in x (0:23m), entre
of gravity iny (0:0m) and airspeed (80 m/s). The RCAM linearised model has
8 states of roll rate (p), pit h rate (q ), yaw rate (r ), roll angle (), pit h
angle ( ) and the inertial body axis velo ities in x (uB ), y (vB ) and z (wB ).
There are 4 ontrol inputs. The a tuator signals are aileron dee tion (A ),
tailplane dee tion (T ), rudder dee tion (R ) and throttle position (T h ).
Additionally, there are 6 wind perturbation inputs of wind velo ity in earth
axis o-ordinates (Wxe , Wye , Wze ) and in body axis o-ordinates (Wxb , Wyb ,
Wzb ). The 7 output measurements used are pit h rate (q), verti al velo ity
(wv ), airspeed (VA ), sideslip ( ), roll rate (p), yaw rate (r ), roll angle ().
3
Program ode for the designs, written using Matlab and Simulink in luding the ontrol
and robust ontrol toolboxes is given in [242.

305
 The a tuator model linearization is ee tively the dynami models of the
a tuators without saturation and rate limits. Port and starboard engines are
ganged together. The linear a tuator transfer fun tion matrix used is
 6:7 6:7 3:35 0:67
diag s+6:7 ; s+6:76 ; s+3:35 ; s+0:67 (21.13)

The nal linearised model for the design has 16 states (8 RCAM, 4 a tuators,
4 delays), 4 ontrol inputs, 6 wind disturban e inputs and 7 sele ted outputs.

System Design Stru ture

The system design stru ture of Figure 21.2 is used. In this system r1 are the
referen e inputs, r2 and r3 are perturbation inputs, z1 are the performan e
outputs, z2 are the weighted ontrol outputs, e1 are the ontroller referen e
inputs, e2 are the fed ba k measurements and u are the a tuator ontrol signals.

z2
W2
r2
+
r1 e1 u + + z1
W1
e2 K r3 G -

Figure 21.2: Two Degree of Freedom RCAM Design System

Minimisation Fun tion

The variables in Figure 21.2 are related by
2 3 2 32 3
z1 W1 M W1 W1 G2 W1 G1 r1
6 z2 7 6 0 0 0 W 2 7 6 r2 7
4 e1 5 = 4
6 7 6 76 7 (21.14)
I 0 0 0 5 4 r3 5
e2 0 I G2 G1 u
Eliminating u, e1 and e2 , the transfer fun tion from r to z is given by
 
W1 (So GK1 M ) W1 So W1 So G2
Tzr = (21.15)
W2 Si K1 W2 K2 So W2 K2 So G2
where the input and output sensitivities are dened as

So = (I + GK2 ) 1 Si = (I + K2G) 1 (21.16)

306
Errors and ontrol eort due to the wind disturban es are minimised via
W1 SoG2 and W2 K2So G2 respe tively.
To minimise Tzr the state spa e stru ture of this plant in orporated into
the design software of [242 is used.

Weight Sele tion

Weight W1 has been sele ted as

 0:7225s 2:5 1 10
W1 = diag s2 +1:7s+0:7225 ; s+ ; s+ ; s+ ;
0:1225s s 10
s2 +0:7s+0:1225 ; s2 +2s+1 ; s+ (21.17)

where  = 10 6 .
The four integral terms are hosen to give zero steady state errors on these
hannels, and will provide good tra king to referen e demands . The integral
gains are a design parameter. In reasing gain gives the losed loop system bet-
ter mat hing to the ideal model and in reases the bandwidth of the disturban e
reje tion.
The longitudinal hannels of verti al velo ity ( wv ) and airspeed (VA ) have
unity gain giving good losed loop performan e and disturban e reje tion. The
 
lateral hannels of sideslip ( ) and roll angle ( ) have gains of 10. These values
are hosen to enable better disturban e reje tion of both the sideslip and roll
angle as well as better tra king the roll angle demands. This is so that errors
on sideslip and roll angle, su h as might o ur during an engine failure, are
qui kly redu ed.
The three rate terms are fed ba k to the ontroller to enhan e ontrol and
stabilisation. The weights on these rates are sele ted to be bandpass lters.
Se ond order lters were sele ted to reje t disturban es and ross- oupling
q
ee ts at the sele ted frequen ies. For the pit h rate ( ), the frequen y sele ted
was taken as the natural frequen y of the verti al velo ity response. For the
p
roll rate ( ), the frequen y sele ted was set to the natural frequen y of the roll
r
angle response, and for the yaw rate ( ) the sele ted frequen y was taken to be
the bandwidth of the heading response ( r
), as yaw rate ( ) is approximately
the same as heading rate.
The ontrol weight W2 needs to limit high frequen y a tivity and to allow
low frequen y tra king. Hen e W2 needs to be a high-pass lter to bound
these requirements. First order weights are sele ted with design parameters of
rossover frequen y and low and high frequen y gains. To redu e the tuning
parameters, the high and low frequen y gains were hosen to be the inverse
of one another. There is no physi al reason why this needs to apply, so more
tuning with an additional parameter ould be applied at a later stage. These
parameters may be more a urately sele ted to represent the un ertainty arising
in the model as this weight an be used to produ e robust designs to plant
additive un ertainty [242. So representing the RCAM entre of gravity and
mass variations as additive un ertainties an help the sele tion riteria for this
weight and lead to robust designs against these variations. The basi weighting

307
fun tion is
ks + w
(21.18)
s + kw
Parameter w is hosen so that high frequen y ontrol is minimised above a
ertain threshold frequen y, and k is hosen to allow low frequen y ontrol
eort and at the same time minimise the high frequen y ontrol eort.
Weights for the a tuators ae ting longitudinal motion, (the tailplane and
engines), were hosen with k = 10 and w = 200 rad/s. Keeping the gain ( ) k
low is desirable so that a tuators operate within a known linear region, in order
to avoid a tuators rate limiting or saturating.
Similarly, for the lateral a tuators (the aileron and rudder), k and w were
hosen as 40 and 200 rad/s. More freedom of ontrol was allowed in the lateral
k
hannels than the longitudinal hannels so the gain ( ) is higher, enabling more
a tuator eort to be used, whi h was parti ularly useful meeting spe i ations
relating to engine failure.
The transfer fun tion of the ontrol eort weight is
 s+200 10s+200 40s+200 10s+200
W2 = diag 40s+8000 ; s+2000 ; s+8000 ; s+2000 (21.19)

Controller Synthesis
The augmented plant is onstru ted by ombining the weights and the lin-
earised model. H1 optimisation is performed and a suboptimal ontroller was
realised with a = 1:1 opt = 5:58. i.e. k Tzr k1 = 5:58, where Tzr is given
in 21.15. This value gives an indi ation of robustness. A more robust solution
would yield a lower value whi h might be obtained at the expense of perfor-
man e for example by redu ing the hannel bandwidths in the weight W1 or
through more tuning of the parameters of ontrol weight W2 for example by
in orporating plant additive perturbations into the sele tion riteria to model
the entre of gravity and mass variations.
H1 optimisation produ es a ontroller whose order will be equal to that
order of the augmented plant. This order is hen e the sum of the number of
states of the plant (8 RCAM + 4 a tuators + 4 delays), weight W1 (10), weight
W2 (4) and the mat hing weight M (6): a total of 36 states. This high order
is redu ed by 10 states using balan ed trun ation.
Some iteration was required to arrive at a reasonable inner loop ontroller.
Having done so, the outer loops were designed.

21.4.2 Outer Loop Design

The inner loop design resulted in a losed loop system that losely approxi-
mated the mat hing model M of 21.11. Therefore, it was de ided to utilise M
as the basi design model for the outer loops design.

Height Tra king

Figure 21.3 shows the loop shaping arrangement for the outer loop height tra k-
ing. The ontroller implementation is shown in Figure 21.4. This arrangement

308
 1
_ 0.7225 wv 1
_ z
Kz
s s2+1.7s+0.7225 s
Wz

Shaped Plant

Figure 21.3: Height Tra king Loop Shaping System

r1 +
1
_ 0.7725 wv 1
_ z
r2 - + Wz Kz
s s2+1.7s+0.7725 s
-

Controller

Figure 21.4: Height Tra king Controller System

was hosen as it an be used to provide good responses to step height demands

(using referen e r1 ) or ramped height demands (using referen es r1 and r2 ),
su h as when following a glide slope.
The shaped plant takes the approximate model for the verti al speed from
the mat hing model 21.11. The verti al speed ( wv ) is integrated to give height
z
( ) as an output. An integrator weight is pla ed at the input to the plant. This
ensures that the system is type 2 and so ensures that there is zero steady state
error to ramped height demands or disturban es. The two outputs of the plant
are shaped using a stati diagonal weight Wz of the form:

 
Wz = kz1 1 0 (21.20)
0 kz 2
Using this loop shaping system, a stabilising sub-optimal ontroller is syn-
thesised using the normalised oprime fa tor method des ribed in [164. Finally,
the ontroller is augmented with the shaping weight to give the stabilising feed-
ba k ontroller.
Design y le iterations are performed, with the weight Wz being tuned. Wz
is tuned by sele ting a nominal k1 and then tuning k2 to meet the required step
response spe i ations. k1 an then be tuned to improve the ramp response,
with minor adjustments to k2 if ne essary. The nal weight hosen was

 
Wz = 0:55 1 0 (21.21)
0 3:64
This led to a = 3:31 whi h is su iently low to indi ate good robustness.
The ontroller generated has 5 states.

309
 1
_ 1 g 1
_
Kh
s s2+2s+1 V s
Wh

Shaped Plant

Figure 21.5: Inertial Tra k Angle Loop Shaping System

r1 +
1_ 1 g
r2 - + Wh Kh
s s2+2s+1 sV
-

Controller

Figure 21.6: Inertial Tra k Angle Controller System

Inertial Tra k Angle Tra king

Figure 21.5 shows the loop shaping arrangement for the outer loop inertial
tra k angle. The ontroller implementation is shown in Figure 21.6. This ar-
rangement is essentially the same as for the height tra king, with two referen e
inputs to allow for step and ramp demands, and an integrator at the input to
ensure zero steady state error to ramped disturban es and referen es.
The shaped plant is based on the approximate model for the roll angle
response from the mat hing model 21.11. The roll angle is integrated and
s aled to give inertial tra k angle as an output. Linearising 21.7 and taking
Lapla e transforms yields:
g
  (21.22)
sV
where g = 9:81 m/s2 and V is the nominal design inertial velo ity. The inertial
tra k angle rate is also made available as an output.
Design y le iterations are then performed as for the height tra king loop.
A stati diagonal weight
 
Wh = 1 0 (21.23)
0 4:15
was hosen. The design realised a = 2:85 indi ating good robustness. The
ontroller generated has 5 states, redu ed to 3 using balan ed trun ation.

Lateral Deviation Reje tion

Figure 21.7 shows the loop shaping arrangement for the outer loop lateral
deviation tra king. The shaped plant takes the losed loop inertial tra k angle
dynami s, resulting from the ontroller Kh designed in the previous subse tion.

310
 _ d +
K W 1 g V
Kh
- s 2+ 2s + 1 sV s
Controller

Inertial Track Angle Loop

Shaped Plant

Figure 21.7: Lateral Deviation Loop Shaping System

In earth oordinates, it is assumed that

Z
Y= V sin dt (21.24)

Assuming the tra k to be along a bearing of 0 deg the lateral deviation (
)

will be equal to Y. The heading deviations when on tra k are assumed to be
small, so   sin . Hen e
V

=  (21.25)
s
where V is the nominal design inertial velo ity.
W
is sele ted to be a simple gain and a stabilising sub-optimal ontroller is
synthesised using the normalised oprime fa tor method. The ontroller order
is then redu ed using balan ed trun ation and the ontroller is augmented with
the shaping weight.
Design y le iterations are then performed, with W
being tuned. The nal
value hosen was
W
= 0:014 (21.26)

whi h realised = 1:85 suggesting good robustness. The ontroller generated

has 7 states due to the high order of the plant model used, but this is redu ed
to 2.

21.4.3 Design Cy le - Dis ussion

Weight Sele tion
Weight sele tion in H1 designs an be a time onsuming pro ess with many
design iterations needed. The RCAM problem was ta kled in su h a way that
weight sele tion was performed e iently. The outer loop design in parti ular
provided three simple H1 ontrollers that ould easily be tuned in a lassi al
manner.
The inner loop has in prin iple more weights to sele t than the outer loops.
However, the mat hing model is dened by the spe i ations, and the weight
W1 is all but dened, ex ept for the integrator gain values. W2 , the ontrol
eort weight, is dened broadly by robustness requirements to be a high pass
lter; only the gains and ut-o frequen ies need to be sele ted, although more
thorough design sele tion might onsider the parameter variations of the RCAM
as additive perturbations and in orporate them a ordingly.

311
 Dening the ontroller ar hite ture an be more time onsuming than the
tuning. Hen e it is important to sele t a good ar hite ture so that the weights
are easily dened.

Controller Order
The design y le presented inherently generates ontrollers with a large number
of states. For the inner loop design, the ontroller will have as many states
as the augmented plant, and for the outer loop, the ontrollers will have more
states than the shaped plants. The redu tion s heme used is a relatively simple
one, introdu ed partly to speed up the performan e tests and simulations.
Further ontroller order redu tion has not been investigated, as it was felt
that the priority was to get su essful designs working. Lower order ontrollers
would be more desirable in an a tual implementation. It is probably possible to
redu e the order signi antly, and so more rigourous ontroller order redu tion
would be re ommended for future work.

Design Time
Overall, the method adopted in this hapter has been found to be very appli-
able to the RCAM problem. Mu h time was needed to nd a suitable system
design ar hite ture, but on e obtained, the produ tion of suitable results was
fairly straightforward with weight sele tion methods as previously dis ussed.
Some time was spent rening and tuning the design.

21.5 Controller Analysis

20 20

0 0

20
20

40
Singular Values [dB]

Singular Values [dB]

40

60
60
80

80
100

100
120

120 140

140 8 6 4 2 0 2 4
160 2 1 0 1 2 3 4
10 10 10 10 10 10 10 10 10 10 10 10 10 10
Frequency [rad/s] Frequency [rad/s]

Figure 21.8: Output Sensitivity and the Sensitivity of the Control Eort to
Input Demands

Frequen y Analysis - Inner Loop

The design seeks to minimise a number of weighted sensitivities as seen in
21.15.

312
 For the output sensitivity of Figure 21.8, low gain is a hieved over the oper-
ating bandwidth for the three mat hed signals of verti al velo ity (wv ), airspeed
(VA ) and roll angle () and the ontrolled sideslip ( ). Pit h (q), roll (p) and
yaw (r ) rates have unity gain at these low frequen ies. At high frequen ies,
the output sensitivity gain is unity and so high frequen y disturban es are not
ontrolled. Between 1 rad/s and 10 rad/s, there are high peaks. Minimising
these peaks in reases the robustness. By the small gain theorem, the higher
the maximum singular value of So , the smaller the unstru tured output inverse
multipli ative perturbation that will de-stabilise the system.

The fun tion Si K1, representing the sensitivity of the ontrol eort to input
demands is also shown in Figure 21.8. At high frequen ies, the gain is small
thus limiting high frequen y ontrol eort as spe ied.

Figure 21.9 shows the frequen y response of the dieren e between the
losed loop system and the mat hing model. At high and low frequen ies
the dieren e is very low. The dieren e is maximised around the operating
bandwidth, whi h may degrade the overshoot or settling time to step demands.
Also shown is the ontroller frequen y response (Figure 21.9).

100 0

50 50
Singular Values [dB]

Singular Values [dB]

0 100

50 150

100 4 2 0 2 4
200 4 2 0 2 4
10 10 10 10 10 10 10 10 10 10
Frequency [rad/s] Frequency [rad/s]

Figure 21.9: Controller Frequen y Response and Dieren e between the Closed
Loop System and the Mat hing Model

Frequen y Analysis - Outer Loop

Figure 21.10 shows the frequen y responses for height tra king. The open
loop model ( Gz ), shaped model (Wz Gz ) and the ontrolled model (Kz Wz Gz )
responses are shown as well as the frequen y response of the losed loop and the
ontroller. Similar analysis is performed for the other outer loop ontrollers.

Controller Eigenvalue Analysis

Figure 21.11 shows the lo ations of the poles of eigenvalues for the stable inner
and ombined outer loop ontrollers.

313
 150 50

100 K
0

50
G 50

0
Gain [dB]

Gain [dB]
100
WG
50

150 GK/(I+GK)
100 KWG

200
150

250
200

250 3 2 1 0 1 2 3
300 3 2 1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10 10 10
Frequency [rad/s] Frequency [rad/s]

Figure 21.10: Frequen y Response of the Height Model, Shaped Plant and
Controlled Plant and for the Height Closed Loop System and the Controller

15 2

1.5
10

5
0.5
Imag Axis

Imag Axis

0 0

0.5
5

10
1.5

15 2
25 20 15 10 5 0 5 4 3 2 1 0 1
Real Axis Real Axis

Figure 21.11: Eigenvalue Lo ations for the Inner and Outer Loop Controllers

Time Domain Analysis

The subsequent analysis shows time domain results of the ontrolled RCAM.
Tables give numeri al results of tests for the nominal system used in the design.
Additionally plots show the linear and non-linear responses for the tests under
dierent ight onditions. The ight onditions used orrespond to those used
in the evaluation s enario.
The ontrolled RCAM has been shown to have good performan e and ro-
bustness at these ight onditions but linear analysis predi ts robustness issues
at some extreme ight onditions. The eigenvalues of the losed inner loop,
onsidered at 81 dierent ight onditions a ross the whole ight envelope,
show inner loop stability. However, when the outer loop is losed, the eigenval-
ues indi ate instability for the ight ondition of extreme aft entre of gravity
and high mass. This maybe a onsequen e of using an experimental design that
used low order `ideal' models, and where the ontrollers were subsequently re-
du ed to simpler low order realisations. Further work on the outer loop design,
in parti ular for extreme ight onditions is needed. Using more a urate de-
sign models and in orporation of perturbation and parameter variations into
the design might be onsidered. Further iterations of the design y le would
be needed in pra ti e.

314
 The ontrol signals are of parti ular interest to monitor to see if ex essive
ontrol eort is being used or if the a tuators are rea hing rate or saturation
limits. For example, the results obtained for the ight path angle are good with
respe t to the spe i ations. However, from the linear simulations, it an be
seen that large and fast tailplane dee tions are alled for. This results in the
non-linear simulation rate limiting, whi h would need to be improved upon.

Rise Time Settling Time Overshoot

tr (se s) ts (se s) Mp (%)
Linear 3.9 20.0 1.5
Non-linear 4.8 16.4 -0.3
Spe i ation < 5.0 < 20.0 < 5.0
Table 21.1: Flight Path Angle Response

6 6
THRUST
4 4
GAMMA
GAMMA
2 THRUST 2
Angles [deg] / Velocities [m/s]

Angles [deg] / Velocities [m/s]

0 0
VCAS VCAS

2 2

4 4

6 6

8 8
DT
10 DT 10

12 12
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
Time [s] Time [s]

Figure 21.12: Linear and Non-Linear Response to a Flight Path Angle Demand

Rise Time Settling Time Overshoot Height Couple

tr (se s) ts (se s) Mp (%) (m)
Linear 11.2 22.0 -0.0 0.0
Non-linear 9.4 37.2 0.0 1.6
Spe i ation < 12.0 < 45.0 < 5.0 < 10.0
Table 21.2: Airspeed Response

315
 15 15

VCAS

Angles [deg] / Velocities [m/s] / Height [m]

Angles [deg] / Velocities [m/s] / Height [m]

10 10
VCAS

THRUST
5 5

THRUST DT
Z
Z
0 0
GAMMA
GAMMA

5 5

DT

10 10
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time [s] Time [s]

Figure 21.13: Linear and Non-Linear Response to an Airspeed Demand

Rise Time Settling Time Overshoot Airspeed Couple

tr (se s) ts (se s) Mp (%) (m/s)
Linear 11.3 23.1 0.0 0.1
Non-linear 11.0 23.2 0.0 0.1
Spe i ation < 12.0 < 45.0 < 5.0 < 0.5
Table 21.3: Height Response

30 30

Z
25 25
Angles [deg] / Velocities [m/s] / Height [m]

Angles [deg] / Velocities [m/s] / Height [m]

20 20

15 15
Z

10 10

DT THRUST
5 5
GAMMA
GAMMA
VCAS
0 0
VCAS
THRUST
5 5

10 10 DT

15 15
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time [s] Time [s]

Figure 21.14: Linear and Non-Linear Response to a Height Demand

Rise Time Settling Time Overshoot

tr (se s) ts (se s) Mp (%)
Linear 9.4 21.3 -0.0
Non-linear 9.4 21.4 -0.0
Spe i ation < 10.0 < 30.0 < 5.0
Table 21.4: Heading Response

316
 15 15

PHI CHI PHI CHI

10 10

5 5
Angles [deg]

Angles [deg]
BETA BETA
0 0

DE DR

5 5

DA DA

10 10
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time [s] Time [s]

Figure 21.15: Linear and Non-Linear Response to a Heading Demand

Heading Rate Roll Angle Roll Angle Overshoot

_ max (deg/s) max (deg) ss (deg) max (%)
Linear 1.3 9.9 3.8 172.0
Non-linear 0.7 8.3 3.7 149.0
Spe i ation < 3.0 < 10.0 < 5.0 < 50.0
Table 21.5: Engine Failure Response

15 15

10 10

5 PHI 5 PHI

CHI CHI
0 0
Angles [deg]

Angles [deg]

BETA BETA

5 DA 5 DA

10 10

15 15
DE DR

20 20

25 25
0 20 40 60 80 100 120 140 160 180 200 0 20 40 60 80 100 120 140 160 180 200
Time [s] Time [s]

Figure 21.16: Linear and Non-Linear Responses to Engine Failure

Rise Time Overshoot

tr (se s) Mp (%)
Linear 28.0 -0.0
Non-linear 27.8 -4.2
Spe i ation < 30.0 < 5.0
Table 21.6: Lateral Deviation Response

317
 10 10

LAT LAT

8 8
Angles [deg] / Deviation [m]

Angles [deg] / Deviation [m]

6 6

4 4

2 2

DA PHI DA PHI
DE DE
0 0
CHI CHI

2 2
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]

Figure 21.17: Lateral Deviation Linear and Non-Linear Responses

21.6 Results of the Automated Evaluation

Methodology-independent results of the ontroller using the evaluation mission
and s enario dened in [145 are now given. Ea h plot shows responses at 4
ight onditions.

Segment I: the ee t of engine failure

As the RCAM represents a twin-engined air raft, a single engine failure will
result in lateral deviation. Figure 21.18 shows the air raft's deviation from the
tra k when an engine fails and is then restarted. The dotted bounds spe ify
the a eptable level of performan e.

First segment: top view Second segment: lateral deviations

300

100
200

50
lateral deviation [m]

100
xdeviation [m]

0 0
0 a b 1
1 c d 2

100
50

200
100

300
20 18 16 14 12 10 8 6 4 2 0 1 2 3 4 5 6 7 8
yposition (YE) [km] along track distance from point 1 [km]

Figure 21.18: Segment I: the ee t Figure 21.19: Segment II: lateral
of engine failure deviations during the 3 deg/s turn

Segment II: the 3 deg/s turn

Figure 21.19 shows the a tual lateral deviations and the a eptable bounds
whilst o-ordinating a 90 deg turn along the tra k.

318
 Segment III: the apture of the -6 and -3 degrees glide slope
Initially, a glide slope of -6 deg is demanded, followed by -3 deg. The verti al
deviations from the desired glide slope and the a eptable bounds are plotted
in gure 21.20.

Third segment: altitude deviations Fourth segment: altitude deviations

30 30

20 20

altitude deviation [m]

altitude deviation [m]

3
10 2 f 3 10
4
g h
0 0

10 e 10

20 20

30 30
16 15 14 13 12 11 11 10 9 8 7 6 5 4 3 2 1
xposition (XE) [km] xposition (XE) [km]

Figure 21.20: Segment III: verti- Figure 21.21: Segment IV: verti-
al deviations from the desired glide al deviations from the desired glide
slope slope

Segment IV: the nal approa h with wind shear

While on nal approa h with a glide slope of -3 deg the ee t of a wind shear
is evaluated. The verti al deviations from the desired glide slope and the a -
eptable bounds are plotted in gure 21.21.

Numeri al results
Table 21.7 gives the numeri al results for the automated evaluation.

Segm. I Segm. II Segm. III Segm. IV Total

Performan e 0.1237 0.1483 0.0545 0.0721 0.0997
Perf. Dev. 0.0249 0.0180 0.0479 0.0833 0.0435
Comfort 0.3674 0.5878 0.5827 0.6235 0.5404
Safety 0.0056 0.0433 0.0077 0.0478 0.0261
Power 0.0030 0.0052 0.0151 0.0303 0.0134

Table 21.7: Numeri al results of the evaluation pro edure

21.7 Con lusions

This hapter applies H1 optimisation to the ontrol of the RCAM.
319
 The ontroller designed onsisted of an inner loop to meet performan e
and robustness requirements during manual piloted ight and an outer loop to
perform the autopilot tra king.
The inner loop was designed using an H1 two degree-of-freedom mixed sen-
sitivity formulation. Performan e spe i ations were translated into method-
dependent obje tives for the verti al velo ity, airspeed and roll angle responses.
Robustness riteria relating to time delays and to wind and output disturban es
as well as to ontrol eort minimisation were all dire tly in orporated into the
design environment. Perturbation and disturban e inputs as well as frequen y
dependent weights were used. Tuning of the weights to meet design require-
ments was needed, but tuning was eliminated where the exa t spe i ations
were in orporated. The dire t in lusion of the design riteria meant that re-
sults produ ed both for ying performan e and robustness analysis were good
at the expense of high order ontrollers.
The outer loop design was simpler in on ept. The inner loop design is
air raft dependent, and mat hes the RCAM response to an ideal air raft re-
sponse. The outer loop design is ee tively air raft independent, making use
of the ideal inner loop response to produ e the autopilot tra king system. H1
loop shaping te hniques generated three outer loop ontrollers to tra k the
height, heading and to reje t lateral tra k deviations. To meet the autopilot
requirements, a number of design iterations to tune the responses were per-
formed, produ ing good tra king results. At extreme ight onditions further
outer loop design iteration is required to improve robustness, possibly at a ost
of some performan e.
The majority of design eort went into reating a suitable system design
ar hite ture. This was very important, as the ar hite ture used allowed many
design spe i ations to be dire tly in orporated. This enabled the time spent
tuning design weights to be redu ed.

320
22. A -Synthesis Approa h (1)

Samir Bennani1 and Gertjan Looye 1

Abstra t. An appli ation of -synthesis to the RCAM design prob-

lem is presented. Our ontrol obje tive is to a hieve robust perfor-
man e. This results in two subproblems that have to be addressed
rst, the stability robustness and the nominal performan e problem.
Under these onditions the stru tured singular value on ept pro-
vides us a tool to a hieve a desired performan e level and keep this
in the presen e of un ertainties. To demonstrate the exibility of
the method separate ontroller ar hite tures have been adopted for
the longitudinal and lateral air raft motions. Further, it is shown
how the spe i RCAM problem requirements are naturally trans-
lated into the design framework. The resulting ontroller performs
satisfa torily and inherits the imposed robustness and performan e
spe i ations.

22.1 Introdu tion

The theoreti al ba kground of -analysis and synthesis is dis ussed in hapter 8.
-Synthesis is a multivariable losed-loop design methodology addressing the
robust performan e issue. We believe that to su essfully apply the method
to the RCAM problem, both a ight me hani s and a lassi al Bode shaping
ba kground are required.
The ight me hani s insight leads to a sensible problem formulation that
is in rst instan e translated into the hoi e of the ontroller ar hite ture.
It is also needed to understand the problem spe i ations, whi h have to be
translated mathemati ally into so alled weighting fun tions. These weighting
fun tions usually ree t (the inverse) frequen y responses of the desired losed-
loop transfer fun tions. With a little Bode shaping ba kground these weighting
fun tions an be onstru ted.
Controller synthesis is automated, so that the major a ent lies on the
weighting fun tion sele tion. This means that the design a tivity has moved
from produ ing a ontroller towards produ ing spe i ations. This in turn
requires the ability to physi ally interpret the requirements.
For the longitudinal motion an energy study has been arried out to analyse
the intera tion between airspeed and ight path. We would like to ontrol these

1
Delft University of Te hnology, fa ulty of Aerospa e Engineering,
Kluyverweg 1, 2629HS Delft, The Netherlands. E-mail: s.bennanilr.tudelft.nl

321
simultaneously through appropriate throttle and elevator ommands. Under-
standing these dynami s provided us guidelines on how to a hieve de oupled
altitude and speed responses with minimum a tuator eort and a pilot like
o-ordination of the ontrols.
For the lateral motion we have studied several trim options for the engine
failure ase [157. This study lead us to opt for the aerodynami ally best
solution: the zero-sideslip option.
In the following se tion we dis uss the hoi e of the ar hite ture. In se tion
3 we show how to ree t the design requirements in a mathemati al format,
suitable for design and analysis: the so alled inter onne tion stru ture. In
se tion 4 we pro eed with the a tual ontroller design where we optimize the
stru tured singular value . In se tions 5 and 6 simulations are presented. We
on lude the work in se tion 7.

22.2 Sele tion of Controller Ar hite ture

The general breakdown of the ontroller ar hite ture is lassi al in the sense
that separate ontrollers for the longitudinal and lateral motions are designed.
The intention is to a hieve the spe i ations as stated in se tion 14.3 without
mode swit hing or other logi . We design for a speed of 80 m/s.
The two main tasks for both ontrollers are stability augmentation (inner-
loop ontrol) and guidan e along a ight path by tra king altitude, lateral
position and airspeed referen es: href , yref and VAref (outer-loop ontrol).

22.2.1 Longitudinal ontroller

The basi ar hite ture of the longitudinal ontroller is depi ted in gure 22.1.
This system uses the tailplane dee tion ( T ) and the throttle positions (T H 1;2 )
to ontrol longitudinal air raft dynami s. The input to the ontroller Klon , on-
sists of attitude angle , pit h rate q (providing for inner-loop bandwidth and
short period damping respe tively) and errors in airspeed and altitude and the
rates thereof. Although  is not available as a measurement, we an estimate
it from other measured signals. V_A is approximated through omplementary
ltering of inertial and air data measurements [148.
The two degrees of freedom lie in the appli ation of a feedba k ontroller
and of feedforward lters, whi h generate smooth altitude and speed om-
mands. These lters are dened by the designer, while the feedba k ontroller
is obtained from optimization via DK-iteration, as des ribed in se tion 8.4.
The losed-loop system has to tra k the ltered altitude and speed refer-
en es as tightly as possible. In fa t, the input lters represent ideal models of
the losed-loop system. This ar hite ture allows the input lters to be repla ed
by any other generator of smooth signals, like the traje tory generator in the
evaluation pro edure, des ribed in se tion 14.3.3.
Note that the feedba k ontroller ombines both inner-loop as well as guid-
an e fun tions.

322
 .
hc hc
h ref feedforward, h
.
VAc VAc
VA ref feedforward, V
.
T h h
.
RCAM VA VA -
K lon TH1,2 (lon)
-
Actuator,
Engines
q

Figure 22.1: Longitudinal ontroller ar hite ture

22.2.2 Lateral ontroller

To ontrol the lateral motions in a o-ordinated way, a dierent ar hite ture is
adopted, see gure 22.2.

ffw
yref c A
y

+
K out
y RCAM
c K in R (lat)
(lat) Actuators

pr

Figure 22.2: Lateral ontroller ar hite ture

Separate inner-loop and outer-loop systems are designed. The inner-loop

ontroller provides stability augmentation and has to tra k roll and sideslip
angle ommands. The inputs to the ontroller are roll and sideslip ommands
and measurements (  ,  , ,  ) and roll and yaw rates (p, r), while the outputs
are aileron and rudder ommands ( A , R ). In most ight ases the sideslip
angle will be regulated to zero ( = 0). Roll angle ommands are provided by
the outer-loop ontroller, whi h has the task of lateral guidan e.

A o-ordinated turn is hara terised by a steady state roll angle if there is no

wind. This ar hite ture allows a roll angle to be ommanded in a feed-forward
path, improving tra king performan e in turns.

The inner-loop ontroller also has a two-degrees of freedom ar hite ture,

sin e full ommands and their orresponding feedba k signals enter the on-
troller instead of error signals. This gives an extra degree of freedom in the
ontroller synthesis (se tion 8.4).

The outer-loop ontroller minimizes lateral ight path errors via roll om-
mands to the inner-loop ontroller. Feedba k of the tra k angle  improves
damping of the lateral outer-loop dynami s.

323
22.3 Translation of Design Criteria into Method
Dependent Obje tives
This se tion des ribes the translation of design riteria into the -framework,
presented in se tion 8.3. We will on entrate on the the longitudinal ontroller
design.

22.3.1 Design riteria in the general framework

u u
.
Act effort
n y nz

(h VA y) ref z w Comfort
W pert
+ u
Nominal Safety
K Act.
Plant
Go
Set of Plants
(h VA y)
other feedback signals

-
Tracking

noise (h VA y) e

Figure 22.3: Implementation of design requirements

In the design spe i ations in se tion 14.3, performan e, safety, a tuator

eort and omfort requirements are distinguished. As we will see, these an
be put in a single performan e ve tor. To ree t the robustness problem the
nominal plant an be extended to a set of plants we design for, using the
perturbation hannels ( , z w, with w =
z ).
In gure 22.3 we see that all performan e variables an be observed by
looking at the signals of interest at their physi al lo ations in the loop. For
example, to wat h a tuator eort, we an pull out the ontrol signals at the
servos. For safety, we an look at the angle of atta k. As a basis for -analysis
as well as synthesis, this representation enables us to take are of nearly all
obje tives simultaneously.
The RCAM problem formulation implemented in the -design framework
an be divided into the following subproblems:

Robust Stability (RS)

The losed-loop system should be stabilized for the spe ied set of plant models.
For example: (input multipli ative un ertainty)

G~ (s) = fG0 (1 + Wpert (s)
) j k
k1  1 g

where G0 represents the linearized air raft and a tuator dynami s at the nom-
inal design onditions,
is a norm bounded (diagonal) matrix and Wpert (s) is

324
a diagonal transfer fun tion matrix shaping the magnitude of the un ertainty
at ea h a tuator input as a fun tion of the frequen y.
The framework allows for implementation of many kinds of un ertainties,
like parametri un ertainties arising in the state spa e matri es, see the design
example in se tion 8.2.

Nominal Performan e (NP)

The losed-loop system should satisfy the performan e spe i ations for the
nominal plant: G0 (s).

Robust Performan e (RP)

This problem goes one step further: the losed-loop system should a hieve the
required performan e level for the whole set of plants we wish to take into
a ount: G~ (s).
Next, we transform the representation in gure 22.3 into the framework
depi ted in g. 22.4. We pull the un ertain element
out of the system, as
well as the sub-system to be designed, the ontroller. To this end we dene two
new sets of inputs and outputs to the nominal plant P.
P in ludes G0 and all losed-loop transfer fun tions indu ed by all input
output signals on the nominal system G0 .

robustness perturbations
tracking Nominal references
act. effort Plant
safety
comfort
P disturbances

Figure 22.4: Design requirements in a general framework

Thus far, we have indi ated the losed-loop quantities of interest in a qual-
itative way. The next step is quanti ation of requirements on them by aug-
menting weighting fun tions.
First, all input signals have to be normalized with respe t to their maximum
value. In other words, we assume that the input signals have magnitude 1 and
need to be s aled to a suitable magnitude (relative to other inputs) before
entering the system. In this way physi al system onditioning is provided.
At the outputs of the performan e hannel we add weighting fun tions spe i-
fying the relative importan e of keeping the signal of interest small as a fun tion
of the frequen y. In a same fashion we add lters to the inputs and/or outputs
of
to shape the magnitudes of the perturbations.

325
 The weighted losed-loop system now looks like:

M = Wout Fl (P; K )Win

where Win and Wout ontain input and output weighting fun tions on the
signals, in luding s alings, performan e weights, perturbation lters, noise-
shaping lters, referen e lters et . This is shown in gure 22.5.
Finally we an apply the general theorems for analysis (se tion 8.3):

1. Robust Stability (M11 (j!)) < 1 8 !

2. Nominal Performan e (M22 (j!)) < 1 8 !
3. Robust Performan e (M (j!)) < 1 8 !
In gure 22.5 the groups of signals involved in ea h of the theorems have been
indi ated (RS, NP, RP).

RS RS
RP Nominal RP
NP Wout Plant Win NP
P

K
M = F l (P,K)

Figure 22.5: Augmentation of weightings

22.3.2 Appli ation to the longitudinal ontroller

In this se tion we will illustrate the prin iples des ribed above with the appli-
ation to the longitudinal ontroller design.
The main performan e obje tive we address here is de oupled tra king of
speed and altitude referen es ( r = [href ; VAref T ), see gure 22.6.
we Wsout Wp e Wid Ws in r
- z
w Wpert
+
Go K
u y=e -

Figure 22.6: Simplied inter onne tion stru ture

The ommands that enter the ontroller are smoothed by a (se ond order,
diagonal) input lter matrix, Wid . We want the losed-loop system to tra k the

326
ltered referen es as lose as possible. This is ree ted by a diagonal weighting
lter matrix Wp , pla ed on the error signals between ideal model response and
the a tual altitude and speed responses ( e = [he ; VAe T ).
In order to normalize the referen e and output signals and to shape the
intera tion level between the altitude and speed hannels, we use input and
output s aling matri es WSin and WSout . Usually, we take WSout = WSin1 .
Un ertainties in the model are ree ted by an input multipli ative un er-
tainty model: (1 +
Wpert ). This way of modeling un ertainties is somewhat
onservative. In a rened design we ould alleviate this by redu ing the mag-
nitude of Wpert and onsidering parametri un ertainties instead, for example
for un ertainty in the mass and the entre of gravity lo ation.

In order to apply the analysis theorems in se tion 22.3.1 we will rst derive
the open-loop inter onne tion stru ture P and obtain the losed-loop inter on-
ne tion stru ture M for a given ontroller K.
Open-loop inter onne tion stru ture P
First, we redraw g. 22.6 into g. 22.7 a ording to the repartitioning depi ted
in g. 22.5 so that the general analysis theorems and synthesis pro edures hold
(se tion 8.3).
The performan e hannels of P are given by the input signals r and the
weighted errors we (we = [whe wVe T ).
We dis onne t the ontroller from the losed-loop stru ture and we write
down the equations for the partitioned open-loop system system P from g. 22.7.

z
w
e
we Wsout Wp Wid Ws in r
-
Wpert
+
Go -
u
y
P
K

Figure 22.7: Transformation to the general system representation

2 3 2 3
z  w 
4 we P11 P12 4 5
5 = r (22.1)
P21 P22
y u
2 32 3
0 0 Wpert w
= 4 WSout Wp G0 WSout Wp Wid WSin WSout Wp G0 54 r 5
G0 Wid WSin G0 u

327
Closed-loop inter onne tion stru ture
To obtain the losed-loop stru ture M = Fl (P; K ) we apply the denition of
the lower linear fra tional transformation:

M = Fl (P; K ) = P11 + P12 K (I P22 K ) 1 P21

With: So = (I + G0 K )
1 , Ti = KG0 (I + KG0 ) 1 [159 and the relation:
K (I + G0 K ) 1 G0 = (I + KG0 ) 1 KG0 , we substitute eq. 22.1 and nd:
 
M= W pert T i W pert KSo Wid WSin
WSout Wp So G0 WSout Wp So Wid WSin
Robust Stability
As we an see from M, robust stability is hara terized by a weighted input
omplementary sensitivity M11 = Wpert Ti . For robust stability we require
(Wpert Ti ) < 1.
Nominal Performan e
The performan e hannel is hara terized by a weighted output sensitivity
WSout Wp So Wid WSin . Now let us on entrate on the spe i ation of de ou-
pled tra king of altitude and speed ommands. The performan e loop in more
detail is:
  " # 
weh Wouth Wph hhe Widh Winh Wouth Wph VhA e WidV WinV href
=
weV WoutV WpV VhAe Widh Winh WoutV WpV VVAe
A
WidV WinV VAref
Usually we take WSout = W
1:
Sin
  " # 
weh Wph hhe Widh Wouth Wph VhAee WidV WinV href
=
weV V
WoutV WpV h Widh Winh WpV VVAe
Ae
Ae
WidV VAref
We an see that with this hoi e of input and output s alings only o-diagonal
terms are ae ted. Taking diag(Winh ; WinV ) = diag(1=10; 10), and diag(Wouth ;
WoutV ) = diag(10; 1=10), the diagonal terms remain un hanged while one o-
diagonal term is amplied and the other is redu ed, both with a fa tor 100.
In this way we an inuen e intera tion between the hannels in ontroller
synthesis.
We an now formulate the nominal de oupled tra king performan e ondi-
tion as: kWSout Wp So Wid WSin k1 < 1.
Robust Performan e
Finally, to obtain the robust performan e ondition for this problem, we require
(M ) < 1. From se tion 8.3.4 we know that this is equivalent to:
 
(M ) = inf  Wpert Ti DWpert KSoWid WSin
D WSout Wp So G0 D 1 WSout Wp So Wid WSin

328
An upperbound for (M ) is: [266

q
(M )  (WSin1 Wid 1 G0 )(kWpert Ti k + kWSout Wp SoWid WSin k)
If WSin Wid = I , the a hievable robust performan e level is proportional to the
square root of the ondition number of the plant ( (G0 )). By properly hoosing
WSin , we an improve the robust performan e level and provide for physi al
system onditioning.

22.4 Controller Design Cy le

22.4.1 Introdu tion
In this se tion the design of the ight ontroller for RCAM is presented.
In se tion 22.2 the ar hite ture of the ontroller is des ribed. Three sub-
systems are designed using -synthesis:
 a longitudinal ontroller, with integrated inner-loop and outer-loop fun -
tions;

 a lateral inner-loop ontroller, for tra king roll and sideslip ommands;

 a lateral outer-loop ontroller, for lateral guidan e.

The design work for these ontrollers onsists of the following steps:

 hoose air raft onguration and ight onditions for the design point
and obtain a linear model for that point;

 dene a lay-out for the general inter onne tion stru ture ( P ), onsisting
of the plant model, ontroller, weighting fun tions, modeled un ertainties;

 iteratively synthesize and analyse ontrollers and adjust weighting fun -

tions. Analysis is based on -plots, ontroller properties (e.g. frequen y
responses, order) and losed-loop time responses;

 preparation of the ontroller for implementation: removal of fast poles,

balan ing, order redu tion;

 implementation with the nonlinear air raft model to perform time simu-
lations.

Of ourse these steps together are also arried out in an iterative way.
The longitudinal ontroller will be des ribed in detail. For the lateral de-
signs the reader is referred to [25. The design work is arried out in Mat-
lab/Simulink with the -Analysis and Synthesis Toolbox (-Tools) [18.

329
22.4.2 Sele tion of the design model
Some basi knowledge of the longitudinal air raft hara teristi s an be ob-
tained from trim urves. These depi t stati (longitudinal) ontrol dee tions,
required for horizontal equilibrium ight, as a fun tion of the airspeed. The
most interesting urves an be found in gure 22.8 (For a more detailed analysis
of the RCAM hara teristi s and performan e, see [157).

a) Trim curves tailplane, varying Xcg b) Trim curves throttles, varying mass
20 15

0.15c
15 m = 120000 kg 150000 kg

DTH1+DTH2 (deg)
0.23c
DT (deg)

10 10
0.31c 120000 kg

5
Xcg = 0.23c
100000 kg
0 5
50 60 70 80 90 100 110 40 50 60 70 80 90 100 110
VA (m/s) VA (m/s)

Figure 22.8: RCAM trim urves

From gure 22.8(a) we an see that for a forward entre of gravity lo ation
the required tailplane dee tion be omes quite large at lower airspeeds. Stati
longitudinal stability is guaranteed, sin e always dT =dVA > 0 [35. More inter-
esting are the throttle trim urves: these show a minimum, whi h o urs at the
minimum drag speed, VMD (whi h depends on air raft onguration and mass).
If the air raft is ying at speed below VMD , it is ying on the ba kside of the
power urve: a lower trim speed requires a higher throttle setting. For more
details, see [35 and [157. For an air raft equipped with separate autopilot and
autothrottle systems, this results in so alled speed instability, when only one
of these systems is engaged [146. This is a strong motivation for a multivari-
able approa h, as is -synthesis, to design an integrated autopilot/autothrottle
system.
We design for a speed of 80 m/s. The nominal design model is obtained
by trimming and linearizing the RCAM in a horizontal ight at a speed of 80
m/s, a mass of 120 000 kg, a verti al CoG position of 0  and a horizontal CoG
lo ation of 0.23 . 
An impression of the model hara teristi s of the model an be obtained by
looking at damping, eigenfrequen ies and time onstants:

mode frequen y (rad/s) damping time onstant

phugoid 0.1269 0.0898 -
short period 1.3837 0.6000 -
Dut h roll 0.6404 0.3684 -
Spiral - - 5.4449
aper. roll - - 0.7682

All modes are stable. The Dut h roll and the short period modes show reason-
able damping while naturally, the phugoid is badly damped.

330
22.4.3 The longitudinal ontroller
As a rst step in -synthesis design, an inter onne tion stru ture is set up. The
stru ture for the longitudinal ontroller is shown in gure 22.9. We will have a

+ h C h C VA CVAC h nom
WS out Wp Wid WS in
- Wact uu VAnom

w z
h h VA VA Wpert +
1
0 RCAM act. 11
00
Wp2 1 (lon)
0
eng. T TH1,2
11
00 K -

noise
q +
nz
q + Wn n
h h VA VA

Figure 22.9: Inter onne tion stru ture for longitudinal ontroller

loser look at ea h of the blo ks in the diagram.

RCAM (lon): This blo k ree ts the linearized longitudinal air raft dynami s.
The general stru ture has been dis ussed in se tion 22.2.1.

a t./eng. : Tailplane a tuator and engine dynami s are modeled as rst order
lters:

HT (s) = 0:151s+1 and HT H1;2 (s) = 1:51s+1

K: This blo k ontains the ontroller to be designed.

Wpert : This diagonal weighting fun tion matrix is a rude way to a ount
for un ertainties in the air raft model, without addressing a spe i un ertain
parameter. With the loop losed via the omplex diagonal matrix
, k
k1  1
a set of models has been dened for whi h we want to guarantee stability and
to a hieve our performan e spe i ations. For
=0 we have our nominal
model.
We assume that the set of models is large enough to a ount for several
un ertainties in the model, like un ertain a tuator responses, un ertain aero-
dynami parameters, time delays, entre of gravity shifts and mass variations.
The -synthesis methodology enables us to address these perturbations as para-
metri un ertainties. This an be done in a rened design.
The weighting fun tions on the diagonal of Wpert have a low value in the lower
frequen y range, while in the mid-frequen y range the magnitude in reases.
Another ee t of this shape is that the ontroller is for ed to roll o at higher
frequen ies (refer to se tion 8.3.6).
Wpert is a diagonal matrix: Wpert = diag(WpertT ; WpertT H1;2 ), with:
s+1
WpertT (s) = 0:25 s=200+1 WpertT H1;2 (s) = 0:25 s=0:200+1
5s+1

In gure 22.10 the frequen y responses of Wpert are given. These weightings
are determined after a few iterations.

331
 5
10

Wp (h,VA)
throttles

mag
0
10 tailplane

Wpert

2 1 0 1 2 3
10 10 10 10 10 10
frequency (rad/s)

Figure 22.10: Frequen y responses of Wp and Wpert

Wid : Wid generates speed and altitude ommands and the rates thereof (se -
tion 22.2.1). The speed and altitude ommands are related to resp. VAref and
href (s) as follows:
h (s) V (s) 0:152
Widh (s) = = A = 2
href (s) VAref (s) s + 2  1  0:15s + 0:152
The lters represent the desired dynami s of the losed-loop system with band-
width and damping as ommonly adopted in autopilot design [132.
Choosing the bandwidthes equal results in de oupled speed and altitude
ontrol with lowest throttle a tivity. This an be explained from point-mass
energy onsiderations. We will give more details in se tion 22.5.1.

Wp : Up to frequen ies beyond the bandwidth of the ideal model lters, the
dieren es between the altitude and speed responses and respe tively h and
VA should be small. This requirement is ree ted by the diagonal weighting
matrix Wp . At lower frequen ies Wp is small, while it rolls o at higher fre-
quen ies. For both the altitude and speed hannels, the weighting fun tions
are taken as

1=50s + 1 2
 
Wph (s) = WpV (s) = 15000 :
1=0:015s + 1
The weights of the rates are set to zero so that the resulting weighting for
the tra king error is given by Wp = diag (Wph ; 0; WpV ; 0). For small steady
state errors the loop gain jGK (j!)j of the system has to be large in the lower
frequen y range. This an be a hieved by in reasing the low frequen y gain of
the these weighting fun tions. By moving the pole to the origin and in reasing
the gain, the ross-over frequen y is held onstant. This is dis ussed in more
detail in [25. In gure 22.10 the frequen y responses of Wp are given. Note
that weighting ontains approximate double integration. From se tion 8.3.6 we
know that the ontroller will at lower frequen ies have approximately the same
shape. This enables the ontrolled system to tra k ramp ommands with very
small steady state error.

332
WSin and WSout : We motivated in se tion 22.3 that we have to normalize the
referen e inputs. h max = 5 m and the
The altitude ommand is s aled to
speed ommand to VA max = 1 m/se , WSin = diag(h max; VA max).
Wp2 : Wp2 = diag(Wpnz ; Wpq ). This weighting is applied in order to keep
ontrol over the pit h rate and over the normal a eleration. Good tra king of
the feedforward lter outputs should not be at the ost of extreme pit h rates.
The weight is onstant over all frequen ies and is set to pq W = 57:3=0:5
; the
pit h rate should never ex eed 0.5 deg/s for the ommand levels adopted in the
inter onne tion stru ture. Normal a eleration nz is weighted with Wpnz =
1=0:02.
Wa t : This weighting fun tion is diagonal and puts onstant weights on the
ontrols and ontrol rates. Wa t is set to:
 
1 ; 1 ;
Wa t = diag T (max 1 ; 1
_ _
) T (max)
T H1;2 (max) T H1;2 (max)
= 57:3 diag 101 ; 151 ; 21 ; 11 ; rad 1
The maximum dee tions and rates are within the limits spe ied in se -
tion 14.2.5.

Wn : Of ourse, measurements will always be orrupted with some noise. Wn

is again diagonal, onsisting of the following weightings:

Wnh = 5  10 4 11==0500
:01s+1
s+1 Wnh_ = 0:02
Wnv = 2:5  10 4 11==0500
:01s+1 W
s+1 nv_ = 0:02
Wn = 0:05=57:3 Wnq = 0:025=57:3

22.4.4 Controller synthesis

Based on the inter onne tion stru ture dened in se tion 22.4.3, we optimize
the performan e robustness index  via DK-iteration (see se tion 8.4). For this
purpose routines are available in -Tools [18.
The -plots are the most important indi ators in the design pro ess: they
give information about whi h weightings need to be adjusted next. For the
longitudinal ontroller the -plot an be found in gure 22.11. The peak value
of  (robust performan e) along the frequen y axis has been minimized via
DK-iteration. Robust performan e is not a hieved: (M ) > 1. Stability is
guaranteed for the whole adopted set of plants (robust stability: (M11 ) < 1).
Nominal performan e is not a hieved in this ase, but onsidering only tra king,
we have met our spe i ations.
The lateral inner-loop and outer-loop ontrollers are designed following the
same pro edure. The resulting ontrollers have high order so that order re-
du tion is ne essary. For this purpose we use balan ed Hankel-singular value
trun ation. Prior to nding a balan ed realization, we have to make sure that
all ontroller poles are stable.
Very fast poles that sometimes o ur, have to be removed before running
simulations. The ontrollers are always he ked for su ient roll o to prevent
the ontrol system from ommanding high-frequen y signals to the a tuators.

333
 Stability/Performance Indicators
1.6

1.4 Robust Performance

1.2

mag
0.8 Nominal Performance

0.6
Nominal Tracking Performance
0.4

0.2
Robust Stability

0
102 101 100 101 102 103
freq rad/sec

Figure 22.11: -plots for the longitudinal ontroller

The ontroller order after redu tion is still high. For the longitudinal system
24, for the lateral inner-loop ontroller 24 and for the lateral outer-loop system
8. Mu h lower orders an be a hieved with more advan ed redu tion methods
[187.

22.5 Analysis of the Resulting Controllers

22.5.1 Longitudinal ontroller
For the longitudinal system we on entrate on tra king performan e and de ou-
pling of the altitude and speed hannels and on minimum ontrol a tivity. An
interesting way to evaluate speed ight path de oupling and throttle a tivity is
to look at energy prin iples. Considering the air raft as a point-mass, we an
write the total energy state (sum of potential and kineti energy) as follows:

1
Etot = Epot + Ekin = mgh + mV 2
2
(we assume there is no wind) Taking the time derivative, dividing by W = mg
and with h_ = V sin , we get the spe i energy rate:
!
V_
E_ s = V sin + (22.2)
g
From equilibrium along the ight path we an derive: [35

T D V_ E_
= + sin = s (22.3)
W g V
Where T and D are resp. thrust and drag. (T D)=W is alled the spe i
ex ess power. Sin e air raft are mostly own near their minimum drag speed
(in that ase dD=dV  0), thrust gives dire t ontrol over the spe i energy
334
rate: an in rease in thrust is fully available for in reasing the ight path angle
and/or the a eleration. The distribution between the two an be ontrolled
by the (energy onservative) elevator.
An interesting way to see whether the ontroller responds a ording to these
prin iples, we an dene an energy ex hange manoeuvre. We ommand altitude
and speed steps so that potential and kineti energy respe tively de rease and
in rease with the same amount (or vi e versa). Sin e we have hosen identi al
feedforward lters for both hannels (see se tion 22.4.3), V_ =g  sin during
the manoeuvre. This means that E_ s  0 and therefore in the ideal ase the
throttles should not respond. Sin e the air raft is not a point-mass, we expe t
some response, but this still should be low. The energy ex hange is ontrolled
with the (energy onservative) tailplane.
The prin iple of using throttles for energy input and elevator for energy
distribution is the basis for the Total Energy Control System (TECS), as
des ribed in [146, 148.

altitude response speed response

0 2.5

2
5
V (m/s)

1.5
h (m)

10
1
15
0.5

20 0
0 20 40 60 0 20 40 60
time (s) time (s)
throttle input Spec. energy rates
8 0.02
contr.
adhoc
vdot/g
6 0.01
DTH1,2 (deg)

Edot_s ()

Edot_s
4 0

2 0.01
sin(gamma)

0 0.02
0 20 40 60 0 20 40 60
time (s) time (s)

Figure 22.12: Energy manoeuvre responses

We now ommand a 20 m altitude hange and a orresponding 2:42 m/s

in rease speed (so that
Ekin =
Epot , starting at VA = 80 m/s). In
gure 22.12 the nonlinear simulation results are plotted. For omparison also
the throttle response is given of an "ad-ho " (in this ontext) design.
In the third sub-gure an be seen that the throttle response is indeed
negligible ompared to the ad-ho  ontroller. In the fourth gure the spe i
energy rate and its ontributions are depi ted. The energy rate remains very
small, while indeed V_ =g  sin .

335
22.5.2 Lateral ontroller
For the lateral design we show responses of the linear model to a lateral posi-
tion step ommand, see gure 22.13a. The outer-loop ontroller produ es a roll
angle ommand, whi h has to be tra ked by the inner-loop ontroller. From
gure 22.13b we an see that this o urs in a very smooth way. The sideslip re-
sponse is small, indi ating su ient de oupling of the roll and sideslip hannels.

12
0.3

10
0.2

PHI, BETA (deg)

8
0.1
Y (m)

6
BETA
0
4
PHI_C PHI

2 0.1

0 0.2
0 20 40 60 80 0 20 40 60 80
time (s) time (s)

a) Lateral step response b) Inner-loop responses

Figure 22.13: Lateral ommand responses

For a more detailed analysis of the ontrol system, we refer to the design
report, [25.

22.6 Results of the Automated Evaluation Pro-

edure
In this se tion we present simulation results and the a hieved performan e of
the omplete ontrol system when implemented in the automated evaluation
mission and s enario, as dened in se tion 14.3.3. Both overall tra king perfor-
man e and inner-loop behaviour of the ontrolled system will be evaluated
by means of bounds on key variables.
We will fo us the dis ussion on the four separate ight segments dened in
se tion 14.3.3. A summary of the numeri al results of the evaluation will be
given at the end of this se tion.

Segment I: the ee t of engine failure

As the RCAM air raft model is twin-engined, a single engine failure initially
results in a lateral deviation. Figure 22.14 provides a top view of the rst
traje tory segment. The given bounds indi ate an a eptable level of perfor-
man e. The ontrol system behaves very well in this situation. We did an extra
simulation without wind disturban es, to obtain insight in how the ontroller
responds to an engine failure, see g. 22.15. The ontrol dee tions are smooth
and the air raft is trimmed to the new situation.

336
 First segment: top view 15

100 10 THROTTLE2

control inputs (deg)

50 THROTTLE1
0
xdeviation [m]

5 DA
0
0 a b 1
10 DT

50
15 DR

20
100

25
20 18 16 14 12 10 8 6 4 2 0 20 40 60 80 100 120
yposition (YE) [km] time (s)

Figure 22.14: Segment I: the ee t Figure 22.15: Segment I: ontrol in-
of engine failure puts during engine failure

Segment II: the 3 deg/s turn

At the beginning of the turn there is a deviation from the referen e traje -
tory; the air raft needs some time to a hieve the bank angle required for a
o-ordinated turn [35. At the end of the turn the opposite o urs: some time
is needed to return to the initial attitude. Figure 22.16 provides a loser look
at the lateral deviations. During the turn, the lateral error is de reased slowly.

Second segment: lateral deviations Third segment: altitude deviations

300 30

200 20
altitude deviation [m]
lateral deviation [m]

100 10 2 f 3

0 0
1 c d 2

100 10 e

200 20

300 30
0 1 2 3 4 5 6 7 8 16 15 14 13 12 11
along track distance from point 1 [km] xposition (XE) [km]

Figure 22.16: Segment II: lateral Figure 22.17: Segment III: verti al
deviations during the 3 deg/s turn deviations from the desired glide-
slope

Segment III: the apture of the -6 and -3 degrees glideslope

The nal approa h starts with a glideslope of -6 deg and after some time this
is in reased to -3 deg. The longitudinal ontroller is implemented without the
feedforward lter for altitude (g. 22.1), sin e the traje tory generator in the
software produ es smooth referen es for altitude ( ). Sin e we required tight h
tra king of the signals from this feedforward lter (se tion 22.4.3), the verti al
deviations from the referen e ight path depi ted in gure 22.17 are very small.

337
 Segment IV: the nal approa h with windshear
Along the glideslope of -3 deg a windshear with downdraft is en ountered.
Equation 22.3 gives the spe i ex ess power for zero-wind onditions. In a
windeld, we would like to onsider the the air raft kineti energy relative
to the airmass. With this in mind, we an a ount for wind disturban es by
orre ting eq. 22.3 with the so- alled F -fa tor [180, 179:
T D V_
= A + sin + F
W g
where F = W_ Xg E os a W_ gZE sin a + WVZAE : WXE and WZE are resp.

horizontal
waV
and verti al wind omponents (in FE ) and a = ar sin VA , ( waV is the

verti al omponent of Va in FV ). A positive F has a performan e de reasing

ee t, sin e a part of the spe i ex ess power (T D)=W is onsumed by
this term. Sin e the airspeed and ight path referen es are onstant, thrust
must hange only to an el F (we assume D  onstant). This is depi ted in
gure 22.18 where indeed
T=W =
(T H 1 + T H 2 )  F .
The windshear en ountered along the ight path learly is not beyond the
performan e limits of RCAM [157. In g. 22.18 we see that thrust is initially
de reased. In severe windshear this is very undesirable, sin e in this phase
the air raft should in rease its energy level in order to y through the ore
of the windeld without loosing too mu h airspeed and altitude. The verti al
deviations from the desired glideslope are plotted in gure 22.19: the deviations
are small sin e the wind disturban e is ompensated by the thrust.

Fourth segment: altitude deviations

30
Ffact. (), THROTTLE 1+2 (rad)

20

0.1 Ffactor
altitude deviation [m]

3
10
4
throttle g h
0

0
10

20

0.1 30
0 10 20 30 40 50 60 70 80 90 100 11 10 9 8 7 6 5 4 3 2 1
time (s) xposition (XE) [km]

Figure 22.18: Segment IV: Figure 22.19: Segment IV: verti al

windshearF -fa tor and throt- deviations from the desired glide-
tle response slope

Numeri al results
Table 22.1 gives numeri al results based on the dis ussed simulation results.
For the motivation and al ulation prin iple of the various numbers see se -
tion 14.3.3.

338
 Segm. I Segm. II Segm. III Segm. IV Total
Performan e 0.1305 0.1626 0.0511 0.0903 0.1086
Perf. Dev. 0.0293 0.0215 0.0371 0.0837 0.0429
Comfort 0.4204 1.1386 1.6027 0.6393 0.9502
Safety 0.0048 0.1038 0.0109 0.0448 0.0411
Power 0.0042 0.0058 0.0152 0.0302 0.0138

Table 22.1: Numeri al results of the evaluation pro edure

22.7 Con lusions

In the previous se tions the appli ation of -synthesis for a ight ontrol system
has been dis ussed. Blo k diagrams depi ting the ontroller ar hite ture an
be found in gure 22.1 and 22.2.
Most eort was put in a hieving good performan e. We designed with gen-
eral (input-multipli ative) un ertainty models, whi h appeared to be su ient
to over variations in mass, enter of gravity lo ation and un ertainty in time
delay. We designed for a speed of 80 m/s. In order to simplify the system we
avoided mode swit hing. Performan e of the designed ontroller is satisfa tory
in terms of tra king and de oupled responses.
A great feature of the methodology is that we an a ount simultaneously for
un ertainties in multiple spe i parameters (e.g. mass, X g ). The  analysis
gives us bounds on the magnitudes of the un ertainties whi h an be tolerated
until instability o urs. The term robustness should in this view be used with
are. The a hieved robustness is in fa e of the addressed un ertainty level.
The plant, ontroller, modeled un ertainties, performan e weights, distur-
ban e models et . all t in a single framework: the general inter onne tion
stru ture (e.g. gure 22.9) and a suitable matrix measure namely the stru -
tured singular value. This framework easily opes with:

 the multivariable nature of the plant;

 dierent kinds of modeled un ertainties;

 simultaneous performan e spe i ations.

As an analysis tool, it enables the designer to assess the performan e and ro-
bustness level of a ontroller in fa e of the modeled un ertainties. On the other
hand, a ontroller an be synthesized from the very same framework. As a
synthesis tool, the framework allows the designer to trade-o between perfor-
man e and robustness obje tives and between dierent performan e obje tives.
Design work onsists of tuning weighting fun tions, while ontroller synthesis
is automated.
A short oming of the method it that the resulting ontrollers are of the order
of the s aled inter onne tion stru ture. Although ommer ially available model
redu tion te hniques give some relief these do not provide satisfa tory redu tion
levels to be a eptable for industrial appli ations as this one. However, mu h

339
eort has been made and en ouraging results have been provided by Pa kard
etal [187. The redu tion problem has been approa hed from a  perspe tive
and in orporated in the original -synthesis problem. The redu tion levels
reported are doubled without loosing the a hieved performan e levels. For our
problem this would mean that it is possible to easily obtain ontroller orders
of six to eight with the proposed te hnique.
As an indi ator for stability,  is generally appli able to assess stability
robustness of a ontrolled plant in fa e of modeled un ertainties. Any (linear)
ontroller designed for this plant an be onne ted in the loop and analysed
[158, 245.
 is a robust performan e indi ator as well. This requires quanti ation be-
tween requirements (e.g. in terms of nonlinear simulations) and weights. From
the design experien e with RCAM we feel that a more fundamental approa h
to this quanti ation needs to be developed.
Beyond , mu h progress has been made in the robust ontrol area. The
potentials of using Linear Matrix Inequalities (LMI's) for ontroller analysis
and synthesis have been widely re ognized. Based on LMI's algorithms have
been developed that an nd robust gain-s heduled ontrollers, taking not only
un ertainties into a ount, but also nonlinearities and the time varying nature
of the plant [260, 13, 112, 257, 185.

340
23. A -Synthesis Approa h (2)

Jan S huring - and Rob M.P. Goverde

1 2 3

Abstra t. Robust ontrol theory deals expli itly with the un er-
tainties in the hara teristi s of the air raft and in the environment
in whi h it has to operate. A modern approa h to robust ontrol is
H1 optimal ontrol. Generally, un ertainty in a model is present
at various omponents and is in this way highly stru tured. The in-
uen e of the stru tured un ertainty on system performan e an be
analysed by applying -analysis. A ombination of the te hniques
of H1 optimal ontrol and -analysis is known as -synthesis. This
method has been applied to the design of an Autopilot for RCAM,
using-synthesis inner-loop ontrollers. The formal riterion of the
-synthesis method ould not be met by these ontrollers. However,
the resulting design performs rather well.

23.1 Introdu tion

Stability and ontrol is one of the major te hni al hallenges in the design of an
air raft. Flight ontrol systems must fun tion properly in all ight onditions
and air raft states in order to ensure safety of ight. The air raft model is a
very omplex system with lots of un ertainties due to unmodelled dynami s,
parameter un ertainty, sensor noise, a tuator errors, et . Also the intera tion
between the air raft and its environment has to be dealt with, in luding dis-
turban es su h as wind shears, gusts, and turbulen e. Furthermore, spe ial
situations su h as engine failure may be taken into a ount.
Besides providing stability and ontrol to air raft, ight ontrol systems
play an in reasingly important role in meeting ost and performan e obje tives
for modern ivil and military air raft. Robust ontrol theory oers a systemati
approa h to design and evaluate ight ontrol systems that fun tion properly
for all relevant ight onditions and air raft states. It deals expli itly with the
un ertainties in the hara teristi s of the air raft and in the environment in
whi h it has to operate.

1
National Aerospa e Laboratory NLR, Anthony Fokkerweg 2, 1059 CM Amsterdam, The
Netherlands. E-mail s huringnlr.nl, fax +31 20 511 3210.
2
The NLR ontribution to GARTEUR A tion Group FM(AG08) was funded by the
Netherlands Agen y for Aerospa e Programs.
3
Delft University of Te hnology, Fa ulteit der Civiele Te hniek, Se tie Verkeerskunde Stev-
inweg 1, Delft, The Netherlands.

341
 A modern approa h to robust ontrol is H1 optimal ontrol theory. This
method an be interpreted as the minimization of the transfer matrix of the
system in a worst ase s enario. The resulting ontrol law guarantees stability
of the losed-loop system and a hievement of the obje tives for the whole lass
of systems indu ed by the un ertainty (i.e. robust performan e), if a ertain
inequality is satised.
Throughout the system, un ertainties may be present. Instead of modelling
these un ertainties as unstru tured as in a worst ase s enario, the un ertainties
at the various omponents an be taken into a ount one by one expli itly, and
rearranged to form a stru tured un ertainty at the system level. A te hnique to
measure robustness, taking expli itly into a ount the stru tured un ertainty,

is stru tured singular value ( ) analysis.
A ombination of the te hniques of H1 optimal ontrol and -analysis
provides an approa h to design and analyse robust ontrol systems. This design
approa h is alled -synthesis. It is des ribed extensively in Chapter 8.
Variations in the behaviour of a system, aused by ertain varying parame-
ters, an be modelled as stru tured un ertainty in the system using parametri
un ertainty modelling, making extensive use of the theory of linear fra tional
transformations.

In the se tions to follow a design proje t is des ribed, illustrating a pro e-

dure to deal with variations in many parameters.
The design will be distributed over a -synthesis longitudinal inner-loop
ontroller, a -synthesis lateral/dire tional inner-loop ontroller and a hand-
tuned outer loop ontroller.
A more detailed des ription of the proje t an be found in [208.

23.2 Control Con ept

23.2.1 Introdu tion
A six degrees-of-freedom non-linear Resear h Civil Air raft Model (RCAM)
in luding a tuators and a model of disturban es have been des ribed in Chap-
ter 14. The air raft has two engines, to be operated in unison.
An Autopilot shall be designed, a ording to the spe i ations in Chap-
ter 14. The Autopilot has to be robust to a number of parameter variations,
a ording to the spe ied operating envelope, listed in Se tion 23.3.2.
Owing to the de oupling property of the air raft dynami s into symmetri
and asymmetri dynami s, the ontrol system design an be divided into lon-
gitudinal and lateral/dire tional ontroller designs. Moreover, a division will
be made into inner and outer loop ontrollers.
The inner loop ontrollers, designed by -synthesis, will provide de oupled
robust ontrol of as many ontrolled variables as there are ontrol ee tors. The
outer loop ontrollers have to provide the link between the spe ied Autopilot
fun tions and the ontrolled air raft.

342
 Owing to the de oupled inner-loop ontrol, the outer-loop ontroller only
has to a t as a number of single-input, single-output ontrollers, redu ing om-
plexity onsiderably. Therefore, its parameters an be tuned by hand to satisfy
the spe i ations.

23.2.2 Longitudinal inner loop ontroller

The following ontrolled variables (equal in number to the number of available
longitudinal ontrol ee tors) have been hosen:

VA airspeed,
wV z omponent of air raft velo ity in FV .
The ontrolled variables shall have responses resembling those of an ideal model
dened in Se tion 23.3. Additionally, the ontroller will use the pit h rate for
damping purposes. This will lessen the burden of the inner loop ontroller to
estimate this quantity.

23.2.3 Lateral/dire tional inner loop ontroller

The following ontrolled variables have been hosen:

 angle of sideslip,
vV y omponent of air raft velo ity in FV .
The variable vV represents the lateral velo ity w.r.t. the ommanded heading.
For an ommanded heading dierent from zero, the a tual lateral velo ity w.r.t.
the a tual ommanded heading has to be al ulated from C , vV and uV .
The ontrolled variables shall have responses resembling those of an ideal
model dened in Se tion 23.3. Additionally, the ontroller will use the roll rate,
the yaw rate and the roll angle, for damping purposes only.

23.2.4 Longitudinal outer loop ontroller

In the outer loop ontroller the ommands for the inner loop are reated. The
ommanded airspeed redu ed by the trim speed, is put through.
The ommanded verti al speed is onstru ted from altitude error, referen e
verti al speed and a tual verti al speed, all provided with appropriate gains.

23.2.5 Lateral/dire tional outer loop ontroller

In the outer loop ontroller the ommands for the inner loop are reated. As no
other requirement for the angle of sideslip has been given than to minimize it,
a zero ommand has been applied. For de rabbing purposes during landing, it
may be used to ontrol air raft heading, to line the air raft up with the runway
dire tion.
The ommanded lateral speed is onstru ted from lateral deviation, ref-
eren e lateral speed, a tual lateral speed (w.r.t. ommanded heading) and
referen e heading rate, all provided with appropriate gains.

343
 parametri un ertainty
modelling

?
- sele tion weights

?
reation
standard plant

?
H1 synthesis 
D -K iteration tuning weights

- -analysis
?
ontroller order
redu tion
?
losed-loop analysis

?
non-linear simulation

?
end

Figure 23.1: The ontrol system design y le.

23.3 RCAM Controller Design

23.3.1 Control system design y le
Figure 23.1 illustrates the generi ontrol system design y le. The rst step
ontains the parameter un ertainty modelling using the theory of LFTs (Chap-
ter 8). Ea h linear system model onsists of four system matri es. For ea h
element of these system matri es, the minimum and the maximum value o -
urring over all operating points onsidered is established. From these values
a mean value and a range for ea h element is al ulated. The elements are
lassied based on mean and range: when both are below a ertain toleran e,
the element is a zero; when only the range is below that toleran e the element
is a onstant; otherwise it is a variable.
Applying the tools from the Parametri Un ertainty Modelling Toolbox [144
a Linear Fra tional Transformation is al ulated from the mean, range and las-

344
perturbation
input
 perturbation
output

--
- air raft

- - -
weighted
noise turbulen e eort
eort
model weighting

ommands
- ideal
model
-+ ? - performan e
weighting
-weighted
errors

ontroller

ontrol measurement

Figure 23.2: Inter onne tion stru ture for ontroller design.

si ation matri es. The LFT represents a omplete set of linear system models.
It ontains the mean model and oe ients, oupling the normalized pertur-
bations into the model. The perturbation matrix is a diagonal matrix with
normalized elements ranging from 1 to 1. An all zero perturbation matrix
will result in the mean model. Any linear model in the respe tive set an be re-
onstru ted from the LFT by a parti ular ombination of perturbation element
values.
In order to apply the -synthesis method an inter onne tion stru ture has
to be reated in the standard plant format, see Figure 23.2. It in orporates both
the air raft LFT and the environment of the air raft. The air raft environment
onsist of disturban es (turbulen e), an ideal model that shall be mimi ked by
the ontrolled air raft, and weightings (eort and performan e weighting). The
omplete standard plant again is formulated as an LFT.
The tuning of weighting fun tions is a manual pro ess, based on notions of
physi al relevan e of the weighting parameters, in view of the way the resulting
time responses of the ontrolled air raft satisfy the ontroller spe i ations.
This tuning pro ess forms the outer loop of Figure 23.1 (not to be onfused
with a ontroller outer loop). It involves a heavy mental eort of the designer.
The -synthesis pro ess is applied to the standard plant. -Synthesis is an
iterative pro ess, hara terized by the D-K iteration. Generally 3 or 4 iterations
are su ient. This iteration forms the inner loop of Figure 23.1. It is a more
or less automated pro ess. Only the D-s aling part of it involves manual

345
intervention.
The resulting ontroller generally has a high order. Moreover, its eigenvalues
may have absolute values that well ex eed pra ti al limits. Therefore order
redu tion and residualisation an be advantageous. Of ourse, the intended
purpose of the ontroller has to be preserved in these operations, in terms of
stability and performan e.
Finally, response al ulations of the ontrolled air raft shall reveal if the
weightings have been hosen properly. If not, the -synthesis pro ess has to be
repeated for an updated standard plant.

23.3.2 Operating envelope

The spe i ation envelope is spanned over a number of operating points dened
by all ombinations of the parameter values given below:

Parameter Nominal Value Lower Bound Upper Bound

time delay 0.075 s 0.05 s 0.1 s
air raft mass 120 000 kg 100 000 kg 150 000 kg
longitudinal CoG 0.23  0.15  0.31 
verti al CoG 0.1  0  0.21 
The spe i ation envelope is used in ontroller design. An impli ation of the
mass sele tion has to be mentioned. Sin e the design airspeed has been dened
as 1.23 Vstall , the design speed depends on air raft mass:
Mass Air speed
100 000 kg 58.1 m/s
120 000 kg 63.7 m/s
150 000 kg 71.2 m/s

23.3.3 Trimming, linearisation and submodel sele tion

In order to apply parametri un ertainty modelling, the non-linear RCAM is
trimmed and linearised in ea h element of an array of 81 operating points,
spanning the operating envelope dened in Se tion 23.3.2. The air raft model
has been augmented by the a tuator dynami s and the time delay (modelled
as a Pad lter to enable linearisation).
From ea h omplete linear model two submodels are sele ted for the longitu-
dinal and the lateral/dire tional dynami s respe tively. This results in two sets
of submodels with mating trimming onditions. These are used in parametri
un ertainty modelling and in linear simulation.

23.3.4 Parametri un ertainty modelling

The method presented applies equally to the longitudinal and the lateral-
/dire tional submodel. The ontents of ea h submodel over the spe i ation
envelope depend on time delay, mass, longitudinal CoG lo ation, and verti al

346
CoG lo ation at the same time. Therefore we have to deal with a 4-dimensional
parametri un ertainty modelling.
A preliminary investigation [94 has shown that ea h varying entry in the
state spa e matri es varies monotonously with respe t to ea h un ertain pa-
rameter, both for separate and for ombined parameter variation. Therefore
the set of values for ea h varying entry only has to ontain a lower and an
upper bound.
From a set of submodels two sets of system matri es are obtained ontaining
the minima and maxima of the entries over all system matri es.
In our approa h the next step is to dene independent variables ea h or-
responding to a varying entry in the system matri es with given lower and
upper bounds. Then, the varying entries are repla ed by the orresponding
independent variables whi h an vary between the lower and upper bound of
the parti ular entries.
In this way the system matri es ontain the whole set of linear models
indu ed by the un ertainty. For the longitudinal submodel 39 entries are las-
sied as varying entries, whi h yields 39 independent variables to des ribe the
un ertainty. Of ourse, the introdu tion of 39 independent variables implies
some onservatism w.r.t. the situation with the 4 original varying parameters.

The varying entries an be seen as a nominal value with a perturbation,

whi h an be regarded as an LFT (Chapter 8). The system an be seen as an
inter onne tion of LFTs whi h an be reformulated as one single LFT, using the
properties of LFTs. The nal perturbation blo k-stru ture for the longitudinal
submodel ontains 39 real perturbation blo ks of unit dimension orresponding
to the 39 varying entries.
The perturbation blo k-stru ture for the lateral/dire tional submodel on-
tains 31 real perturbation blo ks of unit dimension orresponding to 31 varying
entries.
The above pro ess is implemented in MATLAB using routines of the -
Analysis and Synthesis Toolbox (mutools) [18 and the Parametri Un ertainty
Modelling (PUM) Toolbox [144, 238.

23.3.5 Creation of the inter onne tion stru ture

The rst step in the ontroller design y le is to build the inter onne tion
stru ture for ontroller design. This involves dening the inputs and outputs,
the various models and weighting fun tions, the perturbation blo k-stru ture,
and the inter onne tions of the omponents.

Perturbed air raft model

The LFT des ription of the air raft has been obtained in the un ertainty mod-
elling step of the former Se tion. The -synthesis pro edure is appli able to
omplex perturbations only, be ause of the following reason.

347
 In the D-K iteration used in -synthesis, the frequen y dependent D-s alings
have to be tted with stable minimum-phase transfer fun tions. In the present
version of the mutools toolbox this an only be realized for diagonal D-s ale
matri es, whi h implies full omplex blo ks in the perturbation blo k-stru ture.
This is implemented by repla ing the n 1-dimensional real s alar blo ks by
n 1-dimensional omplex blo ks (where n equals 39 or 31 depending on the
submodel). Obviously, a 1-dimensional blo k is always a full blo k. Due to the
repla ement again some onservatism is introdu ed in a subsequent ontroller
design.
Finally, a ontrol performan e perturbation blo k is dened, and the per-
turbation blo k-stru ture is augmented with that blo k in order to express the
robust performan e problem formulation.

Inter onne tion stru ture

Additionally to the perturbed air raft model a number of other blo ks has to be
in orporated into the inter onne tion stru ture, see Figure 23.2. As an example
the longitudinal inter onne tion stru ture will be elaborated.
The turbulen e model has been dened a ording to [145. Turbulen e is a
sto hasti pro ess that an be des ribed by velo ity spe tra. It an be simu-
lated by passing white noise through a lter modelled a ording to a velo ity
spe trum. Commonly used velo ity spe tra for turbulen e modelling are the
Dryden spe tra.
Figure 23.3 shows the frequen y response of the turbulen e lters.

1
Turbulence model : longitudinal (), vertical () 0
Ideal model : q (), wv (), va (.)
10 10

0
10
1
10
1
10

2
10
2
10
3
10

4 3
10 10
2 0 2 4 4 2 0 2 4
10 10 10 10 10 10 10 10 10
Frequency (rad/s) Frequency (rad/s)

Figure 23.3: Frequen y response of Figure 23.4: Frequen y response

longitudinal and verti al turbulen e of ideal model for longitudinal re-
models. sponse.

An ideal model has been dened pres ribing the way the ontrolled air raft
should behave to satisfy the spe i ations. For ea h of the 3 ontrolled output
variables an ideal response is dened. The pit h rate should be minimized for
omfort purposes, so its ideal behaviour is dened as a onstant small gain
of 0:001 . The ideal models for the other ontrolled variables, verti al speed

348
and airspeed are dened by respe tively:

s + 20 s + 15
WwV ideal := 0:01 and WVA ideal := 0:01 :
s + 0:2 s + 0:15
Figure 23.4 shows the frequen y response of the ideal models.

Next, performan e weighting fun tions and eort weighting fun tions have
to be dened. These are the main inputs to the iterative designing pro ess.
The performan e outputs for the ontroller design inter onne tion stru ture
are dened as the weighted errors of the air raft output variables w.r.t. the
ideal behaviour.
The performan e weighting fun tion is an important tool to obtain desired
performan e of the ontroller to be designed. A frequen y dependent weighting
is applied. At high frequen ies the ontroller will not be able to ountera t any
errors due to the inertia of engines and air raft. In the lower frequen ies a
more pre ise ontrol is desired, and in the lowest band, where (approximate)
onstant errors o ur, a perfe t removal is required. A performan e weighting
aimed at su h ontrol hara teristi s behaves like a lag-lead lter:

! s + !1
K 2 ; (23.1)
!1 s + !2
with !1  !2 .
The ontroller resulting from su h a weighting exhibits proportional and
integral a tion. Moreover, the synthesis tools automati ally add derivative
a tion for damping purposes. The integral a tion, evoked by the emphasis on
low frequen y performan e, may ause overshoot. On the other hand, if no
integral a tion is applied, a steady-state error is introdu ed.
In parti ular, the weighting fun tion of the verti al speed error was nally
dened as:
s + 20
WwV := 0:01 ;
s + 0:02
and the weighting fun tion of the airspeed error was dened as:

s + 10
WVA := 0:01 ;
s + 0:01
whereas the weighting of pit h rate has been dened at a onstant 0:001 . Fig-
ure 23.5 shows the performan e weighting fun tions.

The eort weighting fun tion is implemented to weight ontrol eort. Es-
pe ially when saturations and rate limiters are present in the a tuators, are
should be taken not to overload the a tuation, whi h is a main purpose of ef-
fort weighting. The weights are sele ted su h that a ertain de oupling in the
tailplane and throttle ontrols is obtained. The tailplane eort weighting was
nally dened as:
s + 0:2
WT := 0:5 :
s + 10

349
 1
Performance weighting : q (), wv (), va (.) 0
Actuator effort weighting : tailplane (), throttle ()
10 10

0
10
1
10

1
10

2
10
2
10

3 3
10 10
4 2 0 2 4 2 1 0 1 2
10 10 10 10 10 10 10 10 10 10
Frequency (rad/s) Frequency (rad/s)

Figure 23.5: Performan e weighting Figure 23.6: Eort weighting fun -

fun tions for longitudinal ontrol. tions for longitudinal ontrol.

This eort weighting will a t as a lead-lag lter, whi h is of the form of (23.1)
but with !1  !2 . It applies the largest weight at the higher frequen ies. The
engine eort weighting was dened as:

s + 0:02
WT H := 0:5 :
s+2
Also this eort weighting a ts as a lead-lag lter. It applies less weight at the
lowest frequen ies than the tailplane weighting but more at medium frequen-
ies. In this way a ertain preferen e is expressed to use throttle at the lowest
frequen ies, and the tailplane at medium frequen ies. At the highest frequen-
ies the weightings for both a tuators are equally high. Figure 23.6 shows the
eort weighting fun tions.
The weighting fun tions presented above are the result of an iteration pro-
ess. The initial sele tions were based on physi al onsiderations.
Finally, the omponents are onne ted a ording to in Figure 23.2.

23.3.6 D-K Iteration

The D-K iteration, based on the mutools routine dkit, starts by al ulating
a ontroller K using H1 synthesis. The initial D-s alings are assumed to be
unity for all frequen ies. As an example a few Figures are shown from the
longitudinal ontroller design pro ess. Figure 23.7 shows the singular values of
the resulting losed-loop system, in the rst iteration. Next,  is omputed for
the losed-loop system at various frequen y points. Figure 23.7 also shows the
( omplex)  plot.
The se ond iteration starts with tting the D-s alings intera tively. For
ea h full blo k in the perturbation stru ture the D-s aling has to be tted.

As an example, Figure 23.8 shows a rst order t for the third D-s aling.
The se ond iteration pro eeds with a new H1 synthesis for the s aled open-loop
system and with omputing  as before.

350
 SINGULAR VALUE PLOT: CLOSEDLOOP RESPONSE CLOSEDLOOP MU: CONTROLLER #1
30 18

16
25
14

20 12
MAGNITUDE

10

MU
15
8

10 6

4
5
2

0 3 2 1 0 1 2 3
0 3 2 1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10 10 10
FREQUENCY (rad/s) FREQUENCY (rad/s)

Figure 23.7: Singular values (left) and  (right) of the losed-loop system in
the rst iteration.

Iteration 1 2 3
Controller order 17 25 43
Total D-s ale order 0 8 26
29.595 2.927 1.521
peak-  16.565 1.886 1.513

Table 23.1: Summary of the D-K iteration for longitudinal ontroller design.

This pro edure is repeated until no more (signi ant) improvement an be

made. Table 23.1 gives the iteration summary for the longitudinal ontroller
design.
The maximum singular value and  plot of the longitudinal losed-loop
system is given in Figure 23.9. The nal value of  is 1:513, so robust perfor-
man e is not satised for the given weightings and the ( onservative) un er-
tainty model that was used.
Table 23.2 gives the iteration summary for the lateral/dire tional ontroller
design. The nal value of  is 2:329, so again robust performan e is not satised.

Iteration 1 2 3 4
Controller order 17 43 45 45
Total D-s ale order 0 26 28 28
43.055 7.449 4.268 2.693
peak-  9.140 6.279 3.222 2.329

Table 23.2: Summary of the D-K iteration for lateral/dire tional ontroller
design.

351
 FITTING D SCALING #3 of 39, W/ORDER = 1
0
10

1
10
Max. singular value and complex mu of closedloop system
2 1.6
10

3 1.4
10

4
1.2
10 3 2 1 0 1 2 3
10 10 10 10 10 10 10
1) mag data 2) newfit 3) previous DK 1
SCALED TRANSFER FUNCTION: OPTIMAL & RATIONAL

Magnitude
30
0.8

0.6
20

0.4
10
0.2

0 3 2 1 0 1 2 3 0 4
10 10 10 10 10 10 10 10 10
2
10
0
10
2
10
4

1) MU UPPER BND 2) UPPER BND WITH RATIONAL FIT Frequency (rad/s)

Figure 23.8: Example of a rst- Figure 23.9: Maximum singular

order D-s aling t. value and  of longitudinal losed-
loop system.

23.3.7 Controller order redu tion

The ontrollers obtained by D-K iteration both have an order of more than 40.
It is natural that the order of ontrol systems designed by -synthesis is very
high. However, the order of the ontrol system an be redu ed by model order
redu tion te hniques.
For this purpose the Hankel state redu tion te hnique is applied using the
mutools routine hankmr. A redu ed ontroller is reated su h, that its transfer
fun tions are equivalent to those of the original ontroller. Figures 23.10 and
23.11 show the bode plots of the 43th order original longitudinal ontroller
obtained by the D-K iteration and the 16th order redu ed ontroller.
An additional he k on the appli ability of a ontroller order redu tion is the
omputation of  for the losed-loop system with the new (redu ed) ontroller.
In this ase, where  does not satisfy the riterion  < 1 this additional he k
is felt to be less appropriate.
The order of the lateral/dire tional ontroller has been redu ed from 45 to
15. Moreover, from this ontroller a few states with very high natural frequen y
have been removed by residualisation using the mutools routine sresid, be ause
high natural frequen ies are not very favourable for the pra ti al appli ation
of a ontroller.

23.3.8 Linear analysis inner loop ontrollers

The ontrollers have been analysed to he k whether they satisfy the purposes
of the inner loops. Of ourse the losed-loop eigenvalues have to ree t a well
damped behaviour over the envelope. The losed-loop eigenvalues all have a
satisfa tory real part over the spe i ation envelope (Se tion 23.3.2).

352
 4 4
10 10

2 2
10 10
Log Magnitude

Log Magnitude
0 0
10 10

2 2
10 10

4 4
10 4 2 0 2 4
10 4 2 0 2 4
10 10 10 10 10 10 10 10 10 10
Frequency (radians/sec) Frequency (radians/sec)
Reduced controller, order 16, output 1 (original ooo ) Reduced controller, order 16, output 2 (original ooo )
400 300
Phase (degrees)

Phase (degrees)
200 200

0 100

200 0

400 4 2 0 2 4
100 4 2 0 2 4
10 10 10 10 10 10 10 10 10 10
Frequency (radians/sec) Frequency (radians/sec)

Figure 23.10: Original and redu ed Figure 23.11: Original and redu ed
order longitudinal ontroller bode order longitudinal ontroller bode
plots, tailplane ontrol. plots, throttle ontrol.

Another linear analysis is the omputation of losed-loop time responses.

The results from su h analyses are essential in tuning the weighting fun tions
for performan e improvement. Important design riteria are the response to
step inputs and the magnitude of ross- oupling between separate hannels.
In Figure 23.12 these results are presented for the longitudinal inner loop
ontrol, for all 81 parameter ombinations of the spe i ation envelope.

23.3.9 Outer loop design

The ontroller shall provide the ontrol a tions whi h have been spe ied ex-
pli itely in Chapter 14. The longitudinal part of the omplete ontroller will
be des ribed in more detail.
The longitudinal ontroller subsystem is presented in Figure 23.13. The
subsystem ontains both an outer loop ontroller and an inner loop ontroller.
The inner loop ontroller onsists of an ideal model input shaping lter
( lon_idmod) and an a tual ontroller (lon_ tl).
The feedba k signals fed to the inner loop ontroller ea h re eive a separate
treatment. The pit h rate is ltered by a lag-lead lter in order to attenuate
losed-loop gain at high frequen ies. A marginal instability whi h did appear
in non-linear simulation for some parameter ombinations had to be suppressed
by this lter. The verti al velo ity is ltered by a high-pass lter, whi h a ts as
a kind of omplementary lter for the ideal model. The low-frequen y ontent
of the a tual verti al velo ity is fed via the forward hannel, whi h will be dealt
with in the following. The airspeed is redu ed by the trim speed.
The forward hannel of the inner loop ontroller omprises the outer loop
ontroller, produ ing referen e values for the inner loop ontroller. The pit h
rate referen e value is zero, as only two motion variables an be ontrolled
independently using the two longitudinal ontrol ee tors.

353
 wv <> wv cmd wv <> va cmd
1
0

0.8
0.02

0.6
0.04
0.4

0.06
0.2

0.08
0
0 10 20 30 40 0 10 20 30 40
time (s) time (s)

va <> wv cmd va <> va cmd

1
0
0.8
0.02

0.04 0.6

0.06
0.4
0.08

0.1 0.2

0.12
0
0 10 20 30 40 0 10 20 30 40
time (s) time (s)

Figure 23.12: Response of longitudinal losed-loop submodel: verti al velo ity

ontrol, airspeed ontrol and ross- oupling between both.

354
 2 kw1 0
WV_C q cmd

1 + + x = Ax+Bu
kz + Mux + x = Ax+Bu Demux 1
Z_C y = Cx+Du
Rate Limiter y = Cx+Du DT
lon_idmod
kw2 lon_ctl
3 +
2
V_C
v0 THROTTLE1

wv
3
THROTTLE2
q s+40
4s+40
Laglead

wv s
4 Demux Mux
longitudinal s+0.2
z Highpass
measurements
va +

Figure 23.13: Longitudinal subsystem of the designed ontroller.

The referen e value for the verti al velo ity is omposed of the altitude
error, via a gain, and the verti al velo ity outer loop error, via separate gains
for ommanded and a tual verti al velo ity. The total referen e value is fed via
a rate limiter to avoid saturation of the rate limiter in the tailplane a tuator.
The airspeed referen e value again is redu ed by the trim speed.
The outer loop gains have been tuned by hand. Owing to the robust inner
loop ontrollers, simple, onstant outer loops without s heduling, are su ient.
The gains have been dimensioned su h, that satisfa tory responses are obtained
for small outer loop ommands in the rst pla e. Then, when in reasing input
magnitudes aused problems, e.g. due to a tuation non-linearities, the om-
mands fed to the inner loop had to be limited. Finally, the evaluation performed
a ording to Chapter 14.3.3 indi ated the need for onsiderable feedforward.

23.3.10 Resulting ontroller omplexity and implementa-

tion issues
As dis ussed in the previous Se tions, the resulting ontroller is a stable, in-
variant ontroller with fairly high order. No swit hing is applied, the only
non-linearities are a rate limiter and a number of axis transformation fun -
tions.
In an a tual implementation the ontinuous dynami parts have to be trans-
lated to dis rete equivalents. A problem still to be solved is the rather high
bandwidth the ontrollers assume. This may pose a problem in a real-time
environment. The number of states as su h should not be problemati for
state-of-the-art omputing hardware.
The hara ter of this type of bla k box ontrollers, in whi h the fun tion
of individual gains and dynami properties annot be explained to a ertifying

355
authority, might pose a more fundamental problem.

23.4 Non-Linear Assessment of the Designed

Controller
In order to assess whether the ontroller designed in Se tion 23.3 satises the
spe i ations established in Chapter 14, a set of assessment routines has been
written, performing non-linear simulations of RCAM with ontroller. The
assessment produ es standardized results, enabling mutual omparison be-
tween dierent ontrollers. The assessment pro edure is des ribed in Version 3
of [145.
The assessment is able to deal with a number of spe i ations in Se -
tion 14.3.2 of [145. The user an spe ify whi h operating ondition or ombi-
nation of operating onditions shall be applied. A huge number of assessment
results an be obtained this way.
As an illustration the results for lateral deviation ontrol assessment and
altitude response assessment are reprodu ed in Figure 23.14, representing all
81 parameter ombinations of the spe i ation envelope.

23.5 Results of the Automated Evaluation Pro-

edure
The landing s enario onsisting of four segments is des ribed in Chapter 14.3.3.

 Segment I: see Figure 23.15, left side. Engine failure is handled very
well apparently.

 Segment II: see Figure 23.15, right side. The redu tion of lateral error
is too slow to omply with the given bounds.

 Segment III: see Figure 23.16, left side. Verti al error redu tion is not
able to omply in time with the requirements.

 Segment IV: see Figure 23.16, right side. The verti al deviations stay
within the bounds.

 Table 23.3: Most of the numeri al values are below one, whi h means
omplian e with the requirements. The ex eptions are the Comfort req.
in Segment II and the Safety req. in Segment IV.

23.6 Con lusions and Lessons Learned

-Synthesis is dened as an iterative pro ess applying alternately H1 synthesis
and -analysis. An LFT des ription was used to model air raft parametri

356
 Lateral step response Altitude step response
y y_c . [m] h h_c . [m]
1.5 0.5

1 0

0.5 0.5

0 1

0.5 1.5
0 10 20 30 40 0 10 20 30 40
time [s] time [s]
vv vv_c . [m/s] wv wv_c . [m/s]
1 1

0.5 0.5

0 0

0.5 0.5
0 10 20 30 40 0 10 20 30 40
time [s] time [s]
da [deg] dt [deg]
5 1

0.5

0 0

0.5

5 1
0 10 20 30 40 0 10 20 30 40
time [s] time [s]
dr [deg] throttle L R [deg]
0.5
2

0 0

2
0.5
0 10 20 30 40 0 10 20 30 40
time [s] time [s]

Figure 23.14: Assessment of lateral deviation ontrol and altitude response for
all parameter ombinations in the spe i ation envelope.

357
 First segment: top view Second segment: lateral deviations
300

100
200

50

lateral deviation [m]

100
[m]
xdeviation

0 0
0 a b 1
1 c d 2

100
50

200
100

300
20 18 16 14 12 10 8 6 4 2 0 1 2 3 4 5 6 7 8
yposition (YE) [km] along track distance from point 1 [km]

Figure 23.15: Left: segment I - the ee t of engine failure. Right: segment II
- lateral deviations during the 3/s turn.
Third segment: altitude deviations Fourth segment: altitude deviations
30 30

20 20
[m]
altitude deviation [m]

3
10 2 f 3 10
altitude deviation

4
g h
0 0

10 e 10

20 20

30 30
16 15 14 13 12 11 11 10 9 8 7 6 5 4 3 2 1
xposition (XE) [km] xposition (XE) [km]

Figure 23.16: Left: segment III - verti al deviations from the desired glideslope.
Right: segment IV - verti al deviations from the desired glideslope.

un ertainties. A somewhat onservative method was used to deal with the

multi-dimensional aspe t of the un ertainties.

The appli ation of -synthesis to the inner-loop ontroller design of the

RCAM Design Challenge ould formally not su eed in a robust perform-
ing ontroller design (  < 1), neither in the longitudinal, nor in the lat-
eral/dire tional ase. Instead,  values in the order of 2 were obtained.
The outer loop design applied hand-tuning of parameters.
Finally, an apparently rather satisfa tory ontroller has resulted. Only its
behaviour at higher frequen ies and the relation between small input response
and large input response still ould benet from some improvement.

It may be argued, that a less onservative un ertainty modelling pro e-

dure might have resulted in lower  values and maybe in a more satisfa tory

358
 Segm. I Segm. II Segm. III Segm. IV Total
Performan e 0.0552 0.5857 0.5888 0.8875 0.5293
Perf. Dev. 0.0088 0.0363 0.1055 0.1906 0.0853
Comfort 0.4412 2.0553 0.9889 0.5791 1.0161
Safety 0.0032 0.0159 0.0056 4.1336 1.0396
Power 0.0024 0.0033 0.0151 0.0296 0.0126

Table 23.3: Numeri al results of the evaluation pro edure

ontroller. On the other hand, an only slightly more demanding envelope spe -
i ation ould an el su h an improvement readily. After all, it annot be
the intention to re ommend a method whi h is only ee tive for very limited
robustness spe ied.
Therefore one may wonder, if it is desirable from a ontrol performan e
standpoint, to apply a xed gain robust ontroller in air raft ight ontrol,
instead of a ontroller with gain s heduling.
A robust ontroller is not optimized for a single situation, but is the result
of a trade-o between performan e and robustness, yielding an optimum for
a set of operating onditions. Gain s heduling, on the ontrary, uses a tual
information on parameter values to optimize ontroller performan e on line for
any situation, based on existing air raft dynami s knowledge.
A method whi h models parameter variations as un ertainties, as shown in
this Chapter, might not make the most ee tive use of the available data on
urrent air raft state and dynami properties.

A hara teristi of the -synthesis method, and more spe i ally the H1
synthesis part thereof, whi h annot be appre iated, is that the designer has
no dire t inuen e on ertain properties of the ontroller, e.g. its bandwidth.
There is even no dire t inuen e on the stability of the ontroller itself, other
than reje ting it. (Although an unstable ontroller formally might be stabilized
by the plant, su h that the ombination is stable.)
Given the plant LFT des ription, the means the designer has at his disposal,
omprising sele tion of the ontroller top level ar hite ture and the dimension-
ing of weighting fun tions, are of limited power.
Perhaps even more than the extent, to whi h an H1 synthesis ontroller will
be viewed as a bla k box devi e by the ertifying authorities, the H1 synthesis
is therefore experien ed as a bla k box method by the designer, whi h is not
quite appre iated.

359
24. Autopilot Design based on the
Model Following Control Approa h

Holger Duda1 , Gerhard Bouwer1 , J.-Mi hael Baus hat1

and Klaus-Uwe Hahn1
Abstra t. The appli ation of the Model Following Control (MFC)
te hnique to the RCAM autopilot design problem is presented. The
overall ontroller stru ture and the design y le are dis ussed. A
detailed ontroller assessment onrms that the design riteria are
met in the design point and for several additional o-design points,
onsidering varied entre of gravity lo ation, air raft mass, airspeed
and additional time delays. The presented design approa h is in-
sensitive to additional time delays (up to 100 mse ) and entre of
gravity variations. Air raft mass and airspeed variations had small
adverse ee ts on the ontroller performan e. The results of the
automated evaluation pro edure demonstrate that the verti al and
lateral deviations from the tra k are well within the limits for all
investigated ases.

24.1 Introdu tion

In hapter 11 a tutorial review of the applied ontrol design methodology is
given. In this hapter the Model Following Control (MFC) te hnique is applied
to the RCAM autopilot design problem. The already stated feature of the MFC
on ept to separate the main three elements ommand blo k, feedforward and
feedba k ontrollers is utilised. Ea h module an be designed and analysed
separately leading to a straight forward design with a transparent overall
ontroller stru ture.
The design y le and the analysis of the ontroller in terms of the applied
methodology are summarised. The robustness of the MFC design is investi-
gated by nonlinear simulations of typi al manoeuvres in time domain for several
o-design points.

24.2 Overall Controller Stru ture

The overall MFC system stru ture of the RCAM design is presented in gure
24.1. The main three elements of the MFC approa h an be identied: the
1
DLR German Aerospa e Resear h Establishment, Institute of Flight Me hani s, D-38108
Brauns hweig

360
 ommand blo k, the feedforward ontroller and the feedba k ontrollers.
The ommand blo k in ludes linear air raft models in the longitudinal and
lateral axes without intera tions between the two. The following referen e sig-
nals from the traje tory generator are used as inputs to the ommand blo k:
the lateral deviation eyb , the desired heading rate _ r , the desired verti al velo -
ityz_r , the airspeed error VA r VA , the position error zr z and the measured
verti al velo ity wv . The outputs of the ommand blo k are the state ve tor
to be ontrolled x1C = [p ; r ; uB ; wB , its time derivative x_ 1C , and the state
ve tor for de oupling x2C = [q ;  ; vb . Additionally, the states  , p , r , 
and q are utilised as input signals to the feedba k ontroller.

Reference x1C = [pc rc uBc wBc]

Signals .
x1C
eyb Inverse
Aircraft
x1C .
. M1x1 +M2x1 Gain
r Schedule
+
A
. x2C = [qc c vBc] + +
GS T
zr Decoupling + +
Command M3 x 2 R
+ +
Block TH1
+
VAr - Feedforward TH2
c Controller
zr pc -
-
rc -
Feedback
c - Controller
qc -
- -
q
Turn
Correction r
p

wV
z
VA

Figure 24.1: RCAM overall ontroller stru ture

The feedforward ontroller only ontains the matri es M1 to M3 , a ording

to equation (11.8).
The feedba k ontroller ontains several proportional and integral terms
to redu e the error between the ommand blo k states  , p , r ,  , q and
the orresponding air raft states. No additional lters or nonlinear elements
are onsidered for the feedba k path with the ex eption of a turn orre tion
term for the measured pit h rate q. This blo k is required in order to keep the
altitude onstant during a turn. Note that no inner loop airspeed ontrol is
utilised.
Additionally, gain s heduling is used for the ontrol surfa e ommands in

361
order to enhan e the robustness against airspeed variations as has been sug-
gested in [143. No further gain s heduling or adjustments of the ontroller
stru ture are taken into a ount.

24.3 The Translation of RCAM Design Criteria

into Method Dependent Obje tives
For the appli ation of the MFC on ept the RCAM design hallenge has to be
separated into three independent subtasks:
- to design a ommand blo k, whi h meets the design riteria,
- to design a feedforward ontroller whi h in ludes an inverse behaviour of
the plant,
- and to design a feedba k ontroller ope with plant model un ertainties
and disturban es reje tion.
The rst subtask an be a omplished without any referen es to the feed-
forward and feedba k ontroller. A simple (linear or nonlinear) air raft model
with autopilot fun tions has to be designed with respe t to the design require-
ments. Pursuing the idea that numerous autopilot have been designed and
implemented in air raft ight ontrol systems in the past, a relevant autopi-
lot/air raft ombination an be hosen for the ommand blo k for this design
problem. For obvious reasons it was de ided to use an available Classi al Con-
trol approa h to the RCAM problem for the presented on ept [88. Therefore,
with respe t to the translation of the design riteria, the same statements as
summarised in hapter 15 are appli able to the present example.
The design goal for the feedforward and feedba k ontrollers is to for e the
plant to follow ommand blo k states in an optimal way. If the ommand
blo k is hanged for new requirements or any other reason the feedforward
and feedba k ontrollers remain un hanged. They are designed only on e for a
spe i system for all demands generated by the ommand blo k. The design
obje tives for these omponents are dis ussed in se tion 11.5.

24.4 The Des ription of the Design Cy le

The omplete design pro ess is performed within the Matlab/Simulink
TM en-
vironment using the fun tions from the Control System Toolbox ex ept the op-
timisation of the feedba k ontroller. For this subtask an o-line optimisation
based on a quadrati sear h algorithm was used [126.
The design y le steps are presented in gure 24.2: This pro ess in ludes
several iterations regarding the denition of the ommand blo k, the design of
the feedba k ontroller and the he king of the omplete MFC system against
the requirements. It has been shown that mainly the feedba k ontroller had
to be adjusted after he king the omplete system, if the ommand blo k has
been designed arefully, e.g. in luding nonlinear a tuator models.

362
 Definition of a design point for linearisation

Linearisation of the nonlinear aircraft model

Definition of an appropriate controller structure

Determination of the feedforward controller

Definition of the command block based on an available RCAM design

Analysis of the isolated command block

Design of the feedback controller

Analysis of the complete MFC system

Assessment of the MFC system for off-design points

Figure 24.2: RCAM design y le for the MFC approa h

Due to the pure linear approa h in this design example, the determination of
the feedforward ontroller did not require any iterations. Note that also the last
step ( ontroller assessment, se tion 24.5) did not require any iterations, be ause
it has been shown that the ontroller, whi h was developed in the design point,
fullled the design riteria also in the investigated o-design points (one-shot
approa h). The single design steps are dis ussed below:

Design point
The design point has been dened to be approximately in the middle of the
ight envelope (nominal ase of the evaluation pro edure):

m = 120:000 Kg, CG = [0:23; 0; 0, VA = 80 m/se , h = 1000 m.

Linearisation
The linearised air raft model was obtained using the trimr am routine, whi h
is supplied with the RCAM design software. A problem was found in this
linearisation routine: the element of the dynami matrix A(9; 4) was not equal
to zero. This oni ts with the physi al ba kground, be ause there should
be no oupling between the roll angle  and the z omponent of the inertial
velo ity w_ B in the design point. It is thought, that this problem is aused by the
fa t that the gradient for the linearisation is only al ulated in one dire tion.
However, for the determination of the feedforward ontroller this element has
been set to zero: A(9; 4) = 0.

Controller stru ture

Sin e the basi stru ture of a MFC system is xed in prin iple (gure 11.2),
only the input and output signals to the main three elements ( ommand blo k,

363
feedforward and feedba k ontrollers) have to be dened. The input signals to
the feedforward ontroller are summarised a ording to equation (11.5) in the
ve tors x 1C and x2C . The elements of the ve tor x 1C have to be dened, while
x2C automati ally ontains the remaining elements of the omplete air raft
state ve tor x.
Assuming that no asymmetri thrust for yaw ontrol is available, the plant
has four input signals: A , T , R and T H 1 . Hen e, the ve tor x 1C also
has to ontain four elements; these have been dened based on a ontrollability
analysis: x1C = [p ; r ; uB ; wB . Note that the sideslip angle  is not in luded,
therefore, the requirement to minimise the sideslip angle has to be implemented
in the lateral model of the ommand blo k. Regarding the sele tion of the
longitudinal states a further promising alternative would have been to use q
instead of wB . However, the use of wB provided su essful results.
The sele tion of input signals to the feedba k ontroller should be appli able
to ivil transport air raft with onventional ontrol surfa es A , T , R and the
throttle T H 1 .

Feedforward ontroller
The pure air raft dynami s without the a tuator models were onsidered to
determine the feedforward ontroller. It is dened by the matri es M1 to M3
a ording to equation (11.7). The sele tion of the state ve tor to be ontrolled
x1C = [p ; r ; uB ; wB leads to a state ve tor for de oupling x2C , whi h in-
ludes the remaining elements of the omplete state ve tor ( q ,  ,  ,
and vB ). In this example the feedforward ontrol matrix M3 a ording two
equation (11.8) has two zero olumns, therefore, the de oupling state ve tor
was redu ed to x2C = [q ;  ; vB (gure 24.1).

Command blo k
The ommand blo k is separated into a lateral and a longitudinal part without
any oupling, gure 24.3. The stru ture of the autopilot fun tions and the
blo ks of the augmented air raft, whi h ontain linear air raft models in luding
ontrol loops, are summarised in [88.
The augmented air raft blo ks ontain linear air raft models in both axes,
whi h have been obtained from a linearisation of the nonlinear RCAM. In the
lateral axis the augmented air raft represents a roll angle ommand system,
while the inner loops have been designed in order to obtain minimised sideslip
angle during rolling. In the longitudinal axis the augmented air raft model
represents a pit h angle ommand system.
Both augmented air raft models ontain nonlinear a tuator models, whi h
has been shown to be very important regarding the MFC performan e.
The outer loops around the ommand blo k represent typi al autopilot fun -
tions su h as lateral tra k hold, glideslope hold, altitude hold and autothrottle.
Note that the demanded verti al velo ity z_r from the traje tory generator is
only used for mode swit hing between altitude and glideslope hold.

364
 The smoothing lter for the demanded heading rate improves the omfort
during a turn. The blo k  _ to  ontains the following relationship for a
steady state turn [35:

 = tan 1 ( _ V0 =g): (24.1)

.
Lateral track hold pc
.
eyb s rc
-0.0054 Augmented
0.05 s + 1 7.2 c
Lateral Aircraft:
+ + cmd Roll Angle pc
0.001 1/s Command,
+ rc
. +- 30 deg No Sideslip
r 1 . c
tan-1 ( V0/g) Angle
s2 + 1.8s + 1 vBc
.
Smoothing to yc

.
zr
zr - z
0.13 1/s
-
wV Altitude hold
. .
3 -0.022 zr = 0 uBc
- cmd .
P=1 wBc
.
I = 0.056 -0.018 zr > 0 Augmented qc
D=5 Mode Longitudinal c
PID Glideslope hold Aircraft
Select zc
Pitch Angle
VAr - VA P=1 thc Command VAc
I = 0.04 0.2 Engine Model .
zc
D = 2.5
Autothrottle
PID
Autopilot functions Command state
vector generation

Figure 24.3: RCAM ommand blo k stru ture

Analysis of the ommand blo k

The isolated ommand blo k an be analysed against the requirements without
the nonlinear air raft model, the feedforward and feedba k ontrollers. For this
purpose the additional outputs y , z , VA , z_ (broken arrows in gure 24.3)
were used to simulate the losed loop ommand blo k. In omparison to the
original Classi al Control approa h some gains had to be adjusted, for example
in the outer lateral tra k hold (gure 24.3).
Table 24.1 presents the performan e of the isolated ommand blo k in terms
of the design riteria regarding the rise time tr , the settling time ts and the
overshoot Mp for unity step responses in lateral tra k, altitude and airspeed.
Additionally, the airspeed and altitude ross oupling parameter V30 (speed
error peak after a 30 m altitude step demand) and z13 (altitude error peak
after a 13 m/se airspeed step demand) are presented. All requirements are
met for the isolated ommand blo k.

365
 Lateral tra k step Altitude step Airspeed step Cross oupling
tr ts Mp tr ts Mp tr ts Mp V30 z13
(se ) (se ) (%) (se ) (se ) (%) (se ) (se ) (%) (m/se ) (m)
6.4 12.2 1.3 9.2 17.6 1.7 5.1 8.3 0.5 0.2 6.1

Table 24.1: Che king the isolated ommand blo k against the design riteria

Feedba k ontroller
The feedba k ontroller was obtained from an o-line optimisation. A quadrati
sear h algorithm minimises the following ost fun tion by hanging the values
of the feedba k gains [126:

Z t=20se
J= w2 + 602 [15  2 + p2 + q2 + 2 dt: (24.2)
t=0
For every optimisation step, the linearised air raft with the nonlinear a tuator
models and an additional time delay of 150 mse was simulated over 20 se .
Between 0.5 se and 1.0 se the system is disturbed with pit h, roll and yaw
2
a elerations of 2 rad/se . After 10 se simulation time, one engine fails. The
weighting fa tor of 60 relates the error of 1 deg or 1 deg/se to 1 m/se . The
sideslip angle  is weighted 15 times stronger than the remaining angles or
angular rates.
Additionally, up to the engine failure the ontrol surfa e dee tions A , T
and R were weighted with a fa tor of 100 in order to suppress disturban es
with low ontrol a tivity. Initial values for the gains to be optimised were
obtained from the ontrol matrix of the linearised air raft.
The stru ture and the nal gains of the feedba k ontroller are presented
in gure 24.4.

Proportional Integral
state error

output q q

T -1.33 -0.133 Longitudinal

r p

A 0. 0. -6.0 -4.0 0.
Lateral
R -5.1 4.0 0. 0. 4.16

Figure 24.4: RCAM feedba k ontroller: stru ture and gains

The measured pit h rate is orre ted using the following equation, whi h is
ne essary to keep the altitude onstant during a turn [35:

366
 g sin2 
qT = : (24.3)
V os 
Analysis of the omplete MFC system
The omplete MFC system in luding the ommand blo k, the nonlinear air-
raft, the feedforward and feedba k ontrollers was analysed using a spe ial
assessment environment. The following items have been examined:

 Performan e of the ontroller,

 Quality of the inversion,

 Ee ts of turbulen e and engine failure.

The performan e of the ontroller at the design point is presented in table

24.2 onsidering an additional time delay of 50 mse . A performan e is a hieved
similar to the isolated ommand blo k (table 24.1) indi ating that the ontrol
system ensures a more or less exa t following of the ommand blo k.

Lateral tra k step Altitude step Airspeed step Cross oupling

tr ts Mp tr ts Mp tr ts Mp V30 z13
(se ) (se ) (%) (se ) (se ) (%) (se ) (se ) (%) (m/se ) (m)
5.6 10.4 3.3 9.2 17.6 1.9 5.2 12.4 0.7 0.27 4.9

Table 24.2: Che king the omplete nonlinear system against the design riteria
(design point)

The quality of the inversion was investigated by means of a omparison

of the ommand blo k states and the air raft states. Figure 24.5 shows the
time histories of a lateral tra k step response. The step time was one se ond.
The upper two diagrams demonstrate the performan e of the ontroller at the
design point and the orresponding ontrol surfa e dee tions. In the lower two
diagrams the roll and sideslip angles of the air raft are ompared with those of
the ommand blo k. A very good mat h is a hieved.
The investigation of turbulen e ee ts proved that the requirement regard-
ing roll angle was met for the design point. The ee ts of engine failure are
presented in gure 24.6: After 1 se simulation time the left hand engine fails
and after 25 se it is restarted. The maximum lateral deviation after the engine
failure is less than 20 m, whi h is well within the requirement. The ontrol sur-
fa e signals show that the engine failure is mainly ompensated by the rudder.
The roll angle never ex eeds 10 deg and its steady state value is less than 5
deg.
A sideslip angle of about 1 deg built up after the engine failure and restart.
The peak of the lateral a eleration during this manoeuvre is about 0.064 g.

367
 Lateral Step Response (Y (m))
2
1
0
Surface Deflections (solid: DA (deg), broken: DR (deg))
5
0
5
Roll Angle (solid: PHI (deg), broken: PHI_c (deg))
0.5
0
0.5
Sideslip Angle (solid: BETA (deg), broken: BETA_c (deg))
0.02
0
0.02
0 5 10 15 20 25 30 35 40
Time (secs)

Figure 24.5: Lateral step response of the omplete MFC system (design point)

This failure ase demonstrates the hara teristi of the MFC system: the
sideslip angle of the ommand blo k  is about zero during the whole ma-
noeuvre time interval and the dieren e between  and  is redu ed within
se onds. This proves the performan e of the feedba k ontroller.

24.5 Analysis of the Resulting Controller in

Terms of the Applied Methodology
In order to investigate the ontroller robustness against parameter variations
a set of 15 test ases is dened, table 24.3 (page 370). The ases e1 to e4 are
taken from the evaluation pro edure, while the 11 additional ases a1 to a11
have been dened with the following strategy: The ases a1 to a5 are dened in
order to analyse the robustness against mass, verti al entre of gravity lo ation
and airspeed variations. The ase a6 to a11 ontain several variations regarding
mass and airspeed, but all with a most aft entre of gravity lo ation and a time
delay of 100 mse a ording to the re ommendations from [143.
At these test points nonlinear simulations have been performed, whi h are
dis ussed below:

Lateral tra k step responses

Figure 24.7 presents the time histories of lateral tra k step responses for the
dened test ases. The overshoot requirement Mp < 5 % is met for all ases.
Generally, it has been observed that Mp is slightly in reased for the low airspeed
ases, for example 4.2 % for ase a4 against 3.3 % for ase e1 (design point), but
it be omes lear that the ontroller is very robust regarding this requirement.

368
 Lateral Deviation (Y (m))
20 Engine Failure Engine restarted
0
20
Surface Deflections (solid: DA (deg), broken: DR (deg))
20
0
20
Roll Angle (solid: PHI (deg), broken: PHI_c (deg))
10
0
10
Sideslip Angle (solid: BETA (deg), broken: BETA_c (deg))
2
0
2
0 5 10 15 20 25 30 35 40
Time (sec)

Figure 24.6: Ee ts of engine failure (design point)

An os illation in the roll angle is observed for ases a6 and a7, whi h are
hara terised by a low airspeed and a low mass.
The ombination of low airspeed and low mass forms the worst ase for this
type of manoeuvre.

1 m Lateral Step Response

2
Y (m)

0
Roll Angle
0.5
PHI (deg)

0.5
0 5 10 15 20 25 30
Time (sec)

Figure 24.7: RCAM ontroller assessment: lateral tra k step responses

Altitude step responses

Figure 24.8 presents the time histories of 30 m altitude step responses for the
dened test ases. Additionally, unity altitude step responses have been inves-
tigated in order to he k the performan e riteria. The overshoot requirement
Mp < 5 % is met for all ases ex ept a9 (Mp = 5:4 %). The airspeed to altitude
ross oupling requirement is met for all ases. The maximum airspeed error

369
 Mass Air Time Trim
Case (Kg) CGx CGy Speed Delay File Comments
(m/se ) (mse )
e1 120000 0.23 0 80 50 r x0017 Design point
e2 120000 0.23 0 80 100 r x0017 Evaluation ase
e3 120000 0.31 0 80 50 r x0217 Evaluation ase
e4 120000 0.15 0 80 50 r x0117 Evaluation ase

a1 100000 0.23 0 80 50 r x1017 low mass

a2 150000 0.23 0 80 50 r x2017 high mass
a3 120000 0.23 0.21 80 50 r x0027 upper CG
a4 120000 0.23 0 63.7 50 r x0010 low airspeed
a5 120000 0.23 0 90 50 r x0016 high airspeed
a6 100000 0.31 0 58.1 100 r x1210 see [143
a7 100000 0.31 0.21 58.1 100 r x1220 see [143
a8 150000 0.31 0 71.2 100 r x2210 see [143
a9 150000 0.31 0.21 71.2 100 r x2220 see [143
a10 120000 0.31 0 90 100 r x0216 see [143
a11 120000 0.31 0.21 90 100 r x0226 see [143

Table 24.3: Denition of test ases for the ontroller assessment

peak of V30 = 0:49 m/se is again obtained for ase a9. Additionally, a lightly
damped airspeed os illation is present for ases a8 and a9, whi h are the low
airspeed ases with high mass. The upper verti al entre of gravity lo ation
worsens the situation slightly.

The ombination of low airspeed, high mass and upper entre of gravity
lo ation forms the worst ase for this type of manoeuvre.

30 m Altitude Step Response

1000
H (m)

980

960

Airspeed Error
VA VA0 (m/sec)

0.4
0.2
0
0.2
0 5 10 15 20 25 30
Time (sec)

Figure 24.8: RCAM ontroller assessment: altitude step responses

370
Glideslope aptures
Figure 24.9 presents the time histories of 6 deg glideslope aptures for the
dened test ases. Additionally, the airspeed errors are presented indi ating
that two ases are hara terised by a low performan e regarding the airspeed
ontrol after the glideslope apture. These are ases a6 and a7, whi h are
hara terised by a low airspeed and a low mass. Another ase with a poor
airspeed ontrol performan e is the low airspeed ase a4 with nominal mass,
whi h has an airspeed error of about 2 m/se at the end of the simulation time
interval.
The ombination of low airspeed and low mass forms the worst ase for this
type of manoeuvre.

8 m/s Glideslope Capture

1200

1000
H (m)

800

600
Airspeed Error
VA VA0 (m/sec)

10

0
0 5 10 15 20 25 30
Time (sec)

Figure 24.9: RCAM ontroller assessment: glideslope aptures

Airspeed step responses

Figure 24.10 presents the time histories of 13 m/se airspeed step responses for
the dened test ases. In this gure a large overshoot is noti ed for the ases
a6 and a7, whi h are hara terised by a low airspeed and a low mass. However,
for unity airspeed step responses the overshoot requirement is met for all ases.
The airspeed to altitude ross oupling requirement is met for all ases, the
maximum altitude error peak is less than 10 m.
The ombination of low airspeed and low mass forms the worst ase for this
type of manoeuvre.

Engine failures
Figure 24.11 presents the time histories of nonlinear simulations with engine
failures for the dened test ases. The roll angle requirement is met for all
ases. Maximum sideslip angle ex ursions of about 1.1 deg after the engine
failure and about -1.5 deg after the engine is restarted are observed for ases
a8 and a9, whi h are the low airspeed ases with high mass. For these two

371
 13 m/sec Airspeed Step Response

VA VA0 (m/sec)
15

10

0
Altitude Error

1010
H (m)

1005
1000
995
0 5 10 15 20 25 30
Time (sec)

Figure 24.10: RCAM ontroller assessment: airspeed step responses

ases an os illation in the sideslip angle is present indi ating a low dut h roll
damping.
The ombination of low airspeed and high mass forms the worst ase for
this type of manoeuvre.

Roll Angle
10
Engine Failure Engine restarted
PHI (deg)

10
Sideslip Angle

1
BETA (deg)

0 5 10 15 20 25 30
Time (sec)

Figure 24.11: RCAM ontroller assessment: ee ts of engine failures

Turbulen e
Figure 24.12 presents the time histories of nonlinear simulations in turbulen e
for the dened test ases. The requirement regarding the roll angle  < 5 deg
is met for all ases. A maximum verti al load fa tor of about 0.3 g is rea hed
for the high airspeed ases a10 and a11, whi h is only slightly higher than the
maximum value for the design point of 0.25 g.
The performan e requirements regarding overshoot and airspeed to altitude
ross oupling for the dened test ases are summarised in gure 24.13. The
overshoots for step responses of the lateral tra k ( ), the altitude ( ) and the y h
v
airspeed ( ) are presented. The overshoot riterion ( Mp < 5 %) is slightly

372
 Roll Angle
5

PHI (deg)
0

5
Vertical Load Factor
0.5
NZ (g)

0.5
0 5 10 15 20 25 30
Time (sec)

Figure 24.12: RCAM ontroller assessment: ee ts of turbulen e

violated only on e for ase a9 (altitude step). The airspeed to altitude ross
oupling riteria are met for all ases.

These investigations demonstrate the performan e and stability robustness

of the ontroller. Generally, it has been shown that the ee ts of time delays and
entre of gravity deviations are negligible, while airspeed and mass variations
have the most adverse ee ts on the ontroller performan e. However, only one
slight violation of the performan e riteria was found, so that the results are
highly satisfa tory.

Overshoot Criteria
6
5
Mp (%)

4 y
3 h
2 v
1
0
e1 e2 e3 e4 a1 a2 a3 a4 a5 a6 a7 a8 a9 a10 a11
Airspeed Error after 30 m Altitude Step
VA30 (m/s)

.5
.4
.3
.2
.1
0 e1 e2 e3 e4 a1 a2 a3 a4 a5 a6 a7 a8 a9 a10 a11
Altitude Error after 13 m/sec Air Speed Step
10
8
z13 (m)

6
4
2
0 e1 e2 e3 e4 a1 a2 a3 a4 a5 a6 a7 a8 a9 a10 a11
Case

Figure 24.13: RCAM performan e requirements: overshoot and airspeed to

altitude ross oupling

373
 24.6 Results of the Automated Evaluation Pro-
edure
This se tion presents the results of the automated evaluation pro edure dened
in [145: both overall tra king performan e and inner-loop behaviour of the
ontrolled system are evaluated by means of bounds on key variables.

Segment I: the ee t of engine failure

As the air raft model is twin-engined, a single engine failure will mainly result
in yaw a eleration and, therefore, lateral deviation.. Figure 24.14 provides a
top view of the rst traje tory segment. The lateral deviations are well within
the limits. Furthermore, the airspeed should not drop below 1:2  Vstall  61:2
m/se ( m = 120000 Kg). The minimum airspeed during the engine failure is
about 79 m/se in this ase.

First segment: top view Second segment: lateral deviations

300

100
200

50
lateral deviation [m]

100
xdeviation [m]

0 0
0 a b 1
1 c d 2

100
50

200
100

300
20 18 16 14 12 10 8 6 4 2 0 1 2 3 4 5 6 7 8
yposition (YE) [km] along track distance from point 1 [km]

Figure 24.14: Segment I: the ee t Figure 24.15: Segment II: lateral
of engine failure. deviations during the 3 deg/s turn

Segment II: the 3 deg/s turn

At the moment when the turn is initiated, the perfe t following of the required
traje tory and the desire to perform a oordinated turn would imply a sudden
hange in the air raft's roll angle, whi h is only possible with an innitely high
roll rate. Obviously, this is unwanted, su h that deviations from the desired
traje tory at the start (and the end) of the turn are unavoidable. Figure 24.15
provides a look at the a tual lateral deviations, whi h are within the limits.

Segment III: the apture of the -6 and -3 degrees glideslope

A des ent with a glideslope of -6 deg is initiated; again it is unavoidable that
the air raft leaves the desired traje tory. It should return to the traje tory
without overshoot within a period of 30 s. After that, a glideslope of -3 deg
has to be established The verti al deviations from the desired glideslope are
plotted in gure 24.16, whi h are well within the limits.

374
 Third segment: altitude deviations Fourth segment: altitude deviations
30 30

20 20

altitude deviation [m]

altitude deviation [m]

3
10 2 f 3 10
4
g h
0 0

10 e 10

20 20

30 30
16 15 14 13 12 11 11 10 9 8 7 6 5 4 3 2 1
xposition (XE) [km] xposition (XE) [km]

Figure 24.16: Segment III: verti al Figure 24.17: Segment IV: verti al
deviations from the desired glides- deviations from the desired glides-
lope lope

Segment IV: the nal approa h with windshear

During the nal approa h with a glideslope of -3 deg the ee t of a windshear
model is onsidered. The verti al deviations from the desired glideslope are
plotted in gure 24.17, whi h are well within the limits.

Numeri al results
Table 24.4 represents the numeri al results based on the previously dis ussed
simulation out ome. For the motivation and al ulation prin iple of the various
results see [145.

Segm. I Segm. II Segm. III Segm. IV Total

Performan e 0.1058 0.1618 0.1541 0.0686 0.1226
Perf. Dev. 0.0177 0.0185 0.2252 0.2215 0.1207
Comfort 0.3651 1.2611 1.4841 0.6054 0.9289
Safety 0.0042 0.0423 0.0080 0.1049 0.0399
Power 0.0026 0.0083 0.0148 0.0308 0.0141

Table 24.4: Numeri al results of the evaluation pro edure

24.7 Con lusions

The Model Following Control (MFC) te hnique is a very exible tool for ight
ontrol law design. It has been utilised for several years at DLR Institute of
Flight Me hani s in both in-ight simulators ATTAS (Advan ed Te hnology
Testing Air raft System) and ATTHeS (Advan ed Te hnology Testing Heli-
opter System), as well as for dierent ight ontrol system resear h proje ts.
The basi philosophy of this method is to put all available information about
the plant to be ontrolled into the feedforward bran h of the ontrol system.

375
Due to an exa t denition of the desired performan e and the limitations of the
pro ess in the feedforward path, one omes to oni t-free ontrol a tions and,
therefore, to minimum feedba k ontrol a tivity for manoeuvres. This leaves
maximum authority to ope with un ertainties and disturban e reje tion en-
han ing the robustness of the design.
The a urately determined separation of the main three elements ommand
blo k, feedforward and feedba k ontrollers helps to provide a simple and mod-
ular ontrol system ar hite ture (gure 11.2). The expression  ommand blo k,
whi h in ludes the model to be followed, has been introdu ed in order to avoid
misunderstandings regarding the plant, whi h is also often alled model. It is
obvious that any model hara teristi s implemented in the ommand blo k are
limited by the dynami s of the plant to be ontrolled.
In omparison to Classi al Control approa hes the MFC te hnique seems to
require an unne essary additional eort regarding the model denition/design
for the ommand blo k, the design of feedforward and feedba k ontrollers.
But several benets were obtained:

 low feedba k ontrol a tivity in omparison to pure feedba k systems,

 separated design of the main three elements by dierent teams with a

lear sharing of responsibilities,

 saving time and money by implementing qualied ontrol laws in the

ommand blo k of new ight ontrol systems,

 transparent ontroller stru ture allows an easy lo ation of design prob-

lems and re-design for other air raft,

 modules for dierent appli ations an be dened, whi h are appli able to
a omplete air raft family ( ommonality).

Espe ially, the last item demonstrates the power of the MFC on ept, when
looking at a broader spe trum of appli ations. The ommonality of ying
hara teristi s for an air raft family is a protable element onsidering pilot
training and erti ation aspe ts. On e an optimum ommand blo k module is
designed it an be used for a omprehensive variety of dierent air raft, while
only the feedforward and feedba k ontrollers have to be adapted to the new
air raft.
On the other hand, on e the feedforward and feedba k ontrollers are de-
signed for a spe i air raft, the ommand blo k an be hanged for any rea-
sons, su h as dierent tasks and requirements, while the ontroller parts remain
un hanged. A representative example for this te hnique is the in-ight simula-
tion. But this feature an also be applied to operational ight ontrol systems;
the ommand blo k an be hanged for dierent ight ontrol modes, su h
as autopilot or Fly-by-Wire modes. The main onstraint to be onsidered in
this general approa h is that the dynami s of the air raft model implemented
in the ommand blo k are not faster than the dynami s of the air raft to be
ontrolled.

376
 For the RCAM design problem an idealised air raft model with typi al au-
topilot fun tions has been installed in the ommand blo k. The air raft model
has been obtained by a linearisation of the nonlinear RCAM at the design
point. The autopilot fun tions are based on an available Classi al Control ap-
proa h to the RCAM design problem (gure 24.3). Nonlinear a tuator models
are implemented in the ommand blo k, be ause otherwise its outputs may
be too demanding. However, if required for any reasons it is also possible to
implement a omplete nonlinear model in the ommand blo k.
Utilising this available solution, the RCAM design problem was essentially
redu ed to the sele tion of the ontroller stru ture, the determination of the
feedforward ontroller and the optimisation of the feedba k ontroller. The
ontroller stru ture for the ivil transport air raft lass with onventional on-
trol surfa es su h as ailerons, elevator, rudder and throttle levers is xed in
prin iple (gure 24.1).
The separation of the three independent subtasks allows a straight for-
ward design. Assuming, that the designer has a basi knowledge of ight
dynami s, the method is very user-friendly. This statement is supported by
the fa t that the main design work was performed by one person who had
no previous experien es of the MFC on ept. He had only limited knowledge
on ontroller design, but a profound ba kground on Handling Qualities and
Matlab/Simulink
TM appli ations. The eort for the design and assessment of
the RCAM problem in luding preliminary do umentation was less than three
man-months.
The feedforward and feedba k ontrollers are linear with the ex eption of a
turn orre tion for the measured pit h rate, whi h keeps the altitude onstant
during a turn, and a simple gain s heduling for the ontrol surfa e dee tions.
In view of an extension to all ight onditions mainly the ontroller part has to
be extended. Additional design points have to be dened in order to al ulate
a set of feedforward ontroller matri es and feedba k ontroller gains. A gain
s heduling with sele ted blending fun tions an be utilised for the al ulation
of the a tual values depending on the ight ondition or air raft onguration
(mainly ap extension). If required, the feedforward ontroller an be extended
to nonlinear equations, but this an in rease the ontroller omplexity tremen-
dously.
If the ommand blo k has been dened arefully, it an remain un hanged
within the entire ight envelope.
The assessment of the MFC system design for the various parameter hanges
demonstrated adequate ontroller performan e and robustness against param-
eter variations. The design riteria regarding overshoot, airspeed to altitude
ross oupling, rise and settling time were met with only one single ex eption.
Generally, the ases with varied airspeed and mass form the worst ases, while
the ee ts of additional time delays and CG variations were negligible.
The results of the automated evaluation pro edure yield a similar over-
all ontroller performan e ompared to the Classi al Control approa h, whi h
forms the basis for the ommand blo k. It would be interesting to perform a
dire t omparison between the two approa hes, whi h is not yet available.

377
 The omplexity of the overall MFC system is basi ally determined by the
omplexity of the ommand blo k. The feedforward ontroller is dened only
by three matri es. A feedba k ontroller using all signi ant states with pro-
portional and integral terms has been shown to be su ient. The order of
the omplete ontroller is 24 ( ommand blo k: 22, feedforward ontroller: 0,
feedba k ontroller: 2).
The stru ture of the MFC on ept is well suited for error dete tion and on-
trol system re onguration strategies. These aspe ts will gain more importan e
and are of high interest for future developments.
It is proposed to sear h for a unied ontroller stru ture for ivil transport
air raft types. Optimum ommand blo k modules for a omplete air raft family
have to be dened, whi h in lude models of air raft with proven ight ontrol
laws for manual ight (Fly-by-Wire) or existing autopilot fun tions.

378
25. Flight Management Using Predi tive
Control

Mihai Huzmezan1 and Jan M. Ma iejowski 1

Abstra t. This hapter investigates the use of Model Based Pre-

di tive Control ( MBP C ), together with more onventional on-
trol based on H1 loop-shaping, to implement an autopilot for the
RCAM whi h performs ight management in addition to stability
augmentation and guidan e. We believe that MBP C represents a
te hnology whi h should be onsidered in the transition from sta-
bility augmentation to ight management systems.

25.1 Introdu tion

Model Based Predi tive Control (MBPC) a ording to many authors has some
very appealing attributes: Simpli ity  the basi idea of MBP C is fairly
intuitive, and an be understood without advan ed mathemati s; Ri hness 
the ommon elements of MBP C s hemes, su h as models, obje tive fun tions,
predi tion horizons, et , an be tailored to spe i problems; and Pra ti ality
 the usual ombination of linear dynami s and inequality onstraints allows
realisti nonlinearities to be handled [47, 173, 265, 222. See Chapter 12 for a
tutorial presentation of MBP C . Notation in this Chapter is onsistent with
that used in Chapter 12.
Constrained MBP C an remedy some of the drawba ks asso iated with
xed gain ontrollers:

 Disturban es of large amplitude on the plant output might saturate the

a tuators. This ould lead to poor output de oupling and potential loss
of stabilisability in the ase of an unstable open loop plant.

 Respe ting ight envelope onstraints is not straightforward and requires

a priori de isions.

 Pre-lters are usually hosen to give the fastest possible response without
saturating the a tuators during a typi al pilot demand. This implies that
small demands are a hieved in the same time as large demands.

We believe that MBP C is a te hnology whi h should be onsidered in the

transition from stability augmentation to ight management systems. This
hapter summarises our experien es and gives some dis ussion of the potential
of MBP C for ight ontrol. It is based on our experien e of:
1
Cambridge University Engineering Dept, Cambridge CB2 1PZ, England

379
 1. Applying MBP C to the RCAM Design Challenge as a stability augmen-
tation system, see [119.

2. Investigating a ombined MBP C /H1 autopilot for the RCAM model,

see [192.

25.2 Remarks on MBP C for Stability Augmen-

tation
This Se tion is in luded in order to give the reader a avour of the design
y le, ontrol strategy and ar hite ture employed when Model Based Predi tive
Control (MBPC) method was used to a hieve stability augmentation. The level
of su ess a hieved, an overall ritique and some re ommendations for designers
are given. For further details see [119.

The ontroller ar hite ture

The ontroller stru ture we have adopted for the RCAM Design Challenge
was an MBP C ontroller (inner loop) in ombination with a onventional
ontroller (outer loop). The design onsisted of separate design of the inner
and outer loops, for ea h of the longitudinal and lateral hannels. We have
adopted this hybrid stru ture for several reasons, partly te hni al and partly
strategi . The RCAM Design Challenge was a good opportunity to he k strong
and weak points of MBP C . Our de ision to use MBP C in onjun tion with a
onventional outer-loop ontroller allowed us to investigate the limits of MBP C
as a Stability Augmentation System (SAS).

The plant internal model

The model used within the MBP C algorithm to generate the ontrol law is
a linearised representation of the nonlinear RCAM model (des ribed in Chap-
ter 14.2) around an operating point situated in the middle of the ight envelope.

The de oupling into longitudinal and lateral hannels for the air raft model
followed onventional pra ti e on ight ontrol systems for ivil air raft. The
linear internal models for the MBP C ontrol were obtained in two stages.
Firstly a linear model was produ ed using the RCAM nonlinear model, then
a orresponding redu ed model was employed for ea h hannel, based on the
de oupling between hannels and the previous experien e.

This Design Challenge imposed restri tions on the available measurements

(no a tuator states or outputs were assumed to be measurable), therefore the
inner MBP C ontroller does not in lude su h a tuator models as part of its
internal model.

380
Measurement and Referen e Signals
The hoi e of measurement signals depends on the quality of available measure-
ments. Sensor models are not provided and therefore not in luded be ause they
were assumed to be perfe t, whi h makes the usual hoi e of measurement sig-
nals di ult. Like other ontrol te hniques MBP C works better when state
measurements are available. Therefore our approa h to the RCAM Design
Challenge assumes estimation of the state from the nonlinear plant output
measurements using the nonlinear plant equations instead of a onventional
estimator.
In general, the generation of the referen e traje tory for MBP C assumes
some knowledge of the set-point to be followed. For the SAS we assume no
knowledge of the future set-point, therefore we generate the referen e traje tory
as r(k + l) = r(k) for l = 0 : : : N2 . For other appli ations of MBP C this
issue has to be re onsidered be ause the lter whi h generates the referen e
from the set-point an be regarded as another design variable whose ee t is
approximately equivalent to an adjustment of weights, but more transparent
for the designer in some ases.
Sin e the RCAM is a transport air raft, the assumption of a onstant set-
point over the predi tion horizon  20 steps ahead an be maintained be ause
of the slow variation of the ommand from one time step to another. The
ommand is provided either by the pilot or the outer loop ontroller.
For the longitudinal hannel the MBP C ontroller is provided with the
q
following referen e signals: pit h rate ( ), air speed ( Va ), verti al rate (z_ ). The
lateral ontroller has the following referen es provided: side slip ( ) and roll

angle ( ).

The design y le
To summarise, the following steps o ur during the design of an MBP C on-
troller for stability augmentation:

1. Determine the requirements for the losed loop behaviour.

2. Make a hoi e of measurement and referen e signals and produ e the

internal model of the plant.

3. Dene the onstraints related to inputs, rates of hange in the inputs,

outputs and states; onstru t matri es that represent these.

4. Choose appropriate dimensions for ontrol and predi tion horizons.

5. Have an initial hoi e of the ost fun tion weighting matri es.

6. Tune the performan e of the losed loop system by performing losed

loop linear and nonlinear simulations. This involves iteration of steps 3,
4, 5.

7. Perform a stability and robustness analysis using analyti expressions, in

the un onstrained ase, or simulations in the onstrained ase.

381
 First segment: top view Second segment: lateral deviations
300

100
200

50

lateral deviation [m]

100
xdeviation [m]

0 0
0 a b 1
1 c d 2

100
50

200
100

300
20 18 16 14 12 10 8 6 4 2 0 1 2 3 4 5 6 7 8
yposition (YE) [km] along track distance from point 1 [km]
Third segment: altitude deviations Fourth segment: altitude deviations
30 30

20 20

altitude deviation [m]

altitude deviation [m]

3
10 2 f 3 10
4
g h
0 0

10 e 10

20 20

30 30
16 15 14 13 12 11 11 10 9 8 7 6 5 4 3 2 1
xposition (XE) [km] xposition (XE) [km]

Figure 25.1: Results of the automated RCAM Evaluation Pro edure in the ase
of MBP C as a Stability Augmentation System

The ontroller behaviour

The ontroller behaviour is assessed from the perspe tive of performan e, ro-
bustness, ontrol a tivity, passenger omfort and safety. Passing the ontroller
through the Design Challenge Evaluation Pro edure des ribed in Se tion 14.3.3
we are able to laim relatively good operation a ording to the various design
riteria des ribed in Se tion 14.3.2.
Table 25.1 gives numeri al results based on the simulation results. For the
motivation and al ulation prin iple of the various results see Se tion 14.3.3.
In general, performan e requirements were met but, as seen in [119 and Fig-
ure 25.1.

Segm. I Segm. II Segm. III Segm. IV Total

Performan e 0.5153 0.2792 0.1943 0.1138 0.2756
Perf. Dev. 0.0797 0.0675 0.0882 0.0601 0.0739
Comfort 0.6027 1.7584 0.6723 1.2348 1.0671
Safety 0.0081 0.0228 0.0100 0.0250 0.0165
Power 0.0055 0.0077 0.0151 0.0295 0.0145

Table 25.1: Numeri al results of the evaluation pro edure

We onsider that at a high level the stru ture of the MBP C ontroller, as
well as the idea, are simple, but the transparen y of the ontroller is redu ed
by the use of an on-line optimiser. Although the elements of the ontroller,

382
namely the predi tor and the optimiser, orrelate learly with the fun tionality
of the ontrol strategy, as des ribed in Chapter 12, the onstrained optimisation
makes a pre ise task allo ation harder.

25.3 A Combined MBP C /H1 Autopilot for Flight

Management Guidan e and Stabilisation
We have onsidered a ight management system as an alternative appli ation
for the MBP C te hnique going beyond the stability augmentation system [192.
Here a novel ar hite ture is presented and tested in the same framework as that
of the RCAM model. This stru ture provides a robust ontroller, from the point
of view of performan e and stability, maintaining relatively low omplexity and
over oming the negative aspe ts in the role onsidered so far.

ATC
Flight Management System
MBPC

Guidance System
Stability Augmentation System
Hinf

Aircraft
Terrain Map

Figure 25.2: The autopilot ar hite ture

25.3.1 The autopilot ar hite ture

A ight management system must optimise long term obje tives su h as pas-
senger omfort and fuel onsumption over a priori known way-points provided
by air tra ontrol. Moreover, urrent requirements in ight management are
in reasingly on erned with 4D navigation  the ight path is time stamped
and orre ted with the along-tra k error as dened in Se tion 14.3.3. The use
of MBP C as a method to design a ight management system seems desir-
able, as all the above goals an be expressed by the standard MBP C problem
formulation.
Stability augmentation and guidan e systems must be designed by a method
that provides disturban e reje tion at both input and output of the plant, noise
reje tion, robust stability, exibility in spe ifying the bandwidth of the on-
troller, tra king and a ri h variety of analysis tools that will aid erti ation.
For these reasons H1 loop-shaping was used to design the stability augmen-
tation and guidan e system. The stru ture of the H1 loop-shaping ontroller
is shown in Figure 25.5. The reader should onsult the Chapter 7 for the de-
sign pro edure employed. We do not laim that an H1 loop-shaping ontroller
383
 Aircraft Delay
States
Mux Demux
Actuators Aircraft Input Aircraft
Mux Measurements
States Scaling Demux Actuators Actuators
Mux References Hinf Commands from
with Measurements Controller Hinf Controller States
Transformation
Matrix MBPC

Selector
States Mux States

Figure 25.3: The ombined MBP C /H1 ontroller

is the only suitable hoi e, but in the aerospa e eld, where it is appropriate
to address the worst ase signals, the goals on erning guidan e and stability
augmentation are well expressed by the standard H1 loop shaping formulation.
By monitoring the a tuators and plant outputs and imposing onstraints
on their behaviour [173, 137, 199 MBP C an help deal with the drawba ks
asso iated with the xed gain ontrollers. For example, the a tuators an be
used to their limits. MBP C will adjust the referen e to the inner losed loop
H1 ontroller and RCAM air raft) so as to avoid violation of the ight enve-
(
lope and ensure good time response hara teristi s. All the above motivate the
following stru ture shown in Figure 25.3, few explanations being ne essary for
a better understanding. The internal model of the MBP C ontroller ontains
the redu ed model of the air raft, H1 loop-shaping ontroller, a tuators and
the sele tor redu ed from 15 to 7 states. The sele tor denes the dynami
system used for the blending of the altitude and verti al speed. The mean-
ing of the transformation matri es shown in Figure 25.3 is asso iated with the
redu tion of the linear inner-loop model.

The ight management system must not destabilise or an el the ee t of

the inner-loop and must be robust to modelling un ertainty. When employing
this novel ar hite ture the MBP C ontroller, see Figure 25.4, has to be robust
to un ertainty within the inner losed loop bandwidth. Hen e the MBP C on-
troller an be less robust than the H1 ontroller as the un ertainty within the
bandwidth is mu h smaller be ause it has been redu ed by the H1 ontroller.
It is our intention that the MBP C ontroller, shown in Figure 25.4, provide
an optimised referen e to the inner losed loop without interfering with the
stabilisation and guidan e. Note that the MBP C de ision variable,
u(k + l),
depends on the inner loop design. The outputs of the MBP C ontroller (in-
puts to the inner-loop) are the blended altitude and verti al speed (z +  z_ )
and the airspeed (Va ). The referen es are assumed to be provided from a data
base provided by the air tra ontrol (ATC) via Data Link. The estimated
or measured states used within the MBP C belong to the air raft, a tuators,
H1 ontroller and the output mixer.
The ombined ontroller stru ture, shown in Figure 25.3, operates as a
multi-rate system having a high sampling rate ( Ts = 0:01 s) for the H1 on-
troller and a low one, Ts = 1 s, for the MBP C ontroller.

384
 +
Inner Loop ZeroOrder 1/z Time
States Hold5 Unit Delay2 Sum Mux Correction
Mux OPTIMISER + Delta u 1/z u
+
Inner Loop ZeroOrder MBPC Memory Sum1 Unit Delay Inner Loop
Otuputs Hold6 Mux Block Command
Reference
Data Base
Reference Mux MBPC
Measurements

Figure 25.4: The outer MBPC ontroller

Altitude and
Vertical rate Mux
Reference
+
Actuators
Sum Precompensator Input
Airspeed Scaling
Altitude Reference
Measurement Mux References
Mux
Vertical Rate
Measurement Selector Subsystem
Hinf
Controller
Air Speed
Measurement Mux
Measurements

Figure 25.5: The inner H1 ontroller

385
 A single ontroller, shown in Figure 25.5, for longitudinal hannel was de-
signed for both stability augmentation and guidan e even though it is onven-
tional to design an inner loop to provide stability and outer loop for tra king.
The advantage of this stru ture is that the designer obtains insight in how
robustness is traded o for good performan e in altitude following and lateral
deviation minimisation. A tuator states and loop delays to simulate omputa-
tional delays were also used. For further details of the H1 ontroller onsult
the design example in Chapter 7.

Moreover, a single ontroller for both SAS and guidan e system (GS) has
the added advantage of redu ing the number of states of the H1 ontroller.
This be omes very important when employing an MBP C outer loop as the
MBP C uses an internal model of the losed loop for predi tion. The higher
the omplexity of the losed loop the longer the optimisation problem will take
to be solved. Note that the H1 loop-shaping ontroller K1 is pla ed in the
forward loop with no pre-lters.

The ight management role of the MBP C ontroller required us to gener-

ate the traje tory for the MBP C o-line using the generator provided within
the onventional RCAM environment. This generator is not a part of our man-
agement system, be ause we have assumed the traje tory information oming
from a data base given via Data Link by the air tra ontrol (ATC). Then
the MBP C ontroller in a re eding horizon manner will use this information
on-line in order to produ e the optimised referen es for the inner loop.

time Memory Mux

Clock Time
Base
Demux

Aircraft Model Demux

Trajectory Reference
Generator
time Data Base
ZeroOrder Reference
Time Correction Hold

Figure 25.6: The ight manager MBPC traje tory generator

The s heme shown in gure 25.6 was employed to produ e, using the dy-
nami al model of the air raft, the 4D referen e traje tory sampled using the
same sampling rate of the MBP C ontroller. In pra ti e this will be repla ed
by a system whi h interpolates the ATC way-points using the air raft dynami
model.

386
25.3.2 The ight management ontroller design pro e-
dure
The design y le
1. Determine the requirements for the ight management, guidan e and
stabilisation systems behaviour

2. Design an H1 loop-shaping ontroller that will ensure disturban e re-

je tion, noise reje tion and robust stability as des ribed in Chapter 7 by
translating the pertinent requirements into dynami pre and post om-
pensators.

3. Choose the implementation and lo ation of the H1 loop-shaping on-

troller in the inner-loop. Produ e a low order state-spa e model of the
inner-loop that will serve as the MBP C internal model.

4. Dene the onstraints related to inputs, rates of hange of the inputs, out-
puts (ight envelope limits) and states (a tuator limits) and onstru t
matri es that represent these over the ontrol and predi tion horizons.
Choose appropriate values for the MBP C tuning parameters: ontrol
and predi tion horizons and the ost fun tion weighting matri es in or-
der to meet safety, omfort and overall ight management system perfor-
man e requirements.

5. Tune the losed loop ost fun tion parameters via losed loop linear and
nonlinear simulations. This involves iteration of step 4.

6. Perform a stability and robustness analysis using the analyti expressions

in the un onstrained ase or time simulations in the onstrained one.

The translation of design riteria into MBP C method dependent

obje tives
The design riteria for the outer MBP C ontroller involve safety, omfort, on-
trol a tivity and performan e of the overall system. These have to be translated
into the hoi e of several MBP C tuning parameters  ontrol and predi tion
horizons, weighting matri es, sampling time and sometimes even onstraint
boundaries. The hoi e of these parameters is based on several theoreti al re-
sults, but ertain rules of thumb as well, integrated together in a trial-and-error
tuning pro edure.
The safety riteria, ree ting the envelope safeguards, will provide the on-
straint limits upon the variables involved in the on-line optimisation. Table 25.2
gives the output onstraints used and their physi al interpretation.
Having a tuator rates and positions available as states of the air raft model
augmented with models of the a tuators and the H1 ontroller, we are able to
impose onstraints on their behaviour as stated in Table 25.3.

387
 Flight envelope onstraints Minimum value Maximum value
Airspeed (m/s) 51:8  1:05 51:8  2:5
Verti al speed (m/s) 30 +30
Altitude (m) 0 15000

Table 25.2: Flight envelope onstraints as MBP C onstraints

Constrained Variable Name Limits Unit

Tail-plane Dee tion 0:436  T  +0:174 rad

Tail Plane Dee tion Rate 0:261  _T  +0:261 rad/s

Engines Throttle Limits +0:009  T h  +0:174 rad

Engine Throttle Slew Rates 0:028  T_ h  +0:028 rad/s

Table 25.3: A tuator onstraints as MBP C onstraints

The way to translate requirements upon omfort within MBP C is by em-

ploying ost fun tion weights that give suitable trade-o between tight following
of a given traje tory and large loads on the air raft stru ture.
The robustness of the ombined MBP C /H1 autopilot is obtained by ap-
propriate design of the inner H1 loop shaping ontroller taking advantage of
its main feature.
Now we larify how the performan e riteria of the MBP C ight manager
are translated into the available tuning parameters. The main performan e
variables that an be measured by using a step as the referen e traje tory are
the rise time (tr ), the settling time ( ts ) and the overshoot (Mp ), as dened
in [82, 159 for example.

The tuning of the ight management MBP C parameters

In the following se tion, where we use the same notation for the tuning param-
eters as in Chapter 12, we give all the te hni alities of the design y le step
involving tuning of the MBP C parameters.

The ontrol Nu and predi tion N2 horizons. The inuen e of the ontrol and
predi tion horizons is mainly upon the performan e of the ontrolled system,
but they have some inuen e upon robustness as well. In general a smaller
ontrol horizon makes the ontrolled system more robust to un ertainties su h
as parameter variations [222. In general the hoi e of these horizons takes into
a ount knowledge of the dynami s of the inner-loop. In ase of the ontrol
horizon we perform a step response analysis of the system assuming a predened
sampling period. Our nal hoi e for the ontrol horizon was Nu =4 after
in reasing it from the minimum value of 1.
The predi tion horizon is derived from the settling time having in gen-
eral a lenght greater than the system order. A small horizon will redu e the

388
 omputational omplexity, but must ontain at least the non-minimum phase
behaviour in the ase of su h systems to be ontrolled using MBP C . The
robust performan e and stability impose extra boundaries on the horizons.
An in rease of the predi tion horizon should be onsidered only if the ontrol
system proves to have long settling time in order to avoid de reasing the speed
of the ontrol algorithm to an una eptable extent.
The ee t of the predi tion horizon upon step responses (via its dening
parameters: rising and settling time) was studied. The inuen e upon settling
time is relatively large for a predi tion horizon from 1 to 8, but signi antly
smaller for horizons greater than 8. For our ost fun tion trading o ompu-
tational omplexity against robust stability we have in reased the predi tion
horizon from 7 (the MBP C internal model order) to N2 =10.
The small ontrol and predi tion horizons ensured that the optimisation
an be solved in real time ( 0:9 s on a Sun SPARCstation 20).
The sampling period Ts . The sampling period plays an important role in
MBP C ontrollers. A possible hoi e for this parameter is ten times smaller
than the fastest settling time in the losed loop system (the value of Ts is ob-
tained using linear time response analysis assuming onstraints are ina tive).
There is a tradeo between de reasing the sampling period and in reasing
dimensions of matri es involved in the MBP C algorithm or the number of
optimisations performed in the time unit. If the sample period is small in
relation with the servo performan e of the system (e.g. settling time) this
will result in large ontrol and predi tion horizons, possibly ausing numeri al
problems. However, the smaller the sampling period, the better an a referen e
traje tory be tra ked or disturban e reje ted.
In order not to interfa e with the inner ontroller, the MBP C ontrol loop
should have a lower bandwidth than the inner loop. This allows a big value
for the sampling time Ts = 1 s, but does not require it. This makes possible
a real time implementation of the ontroller and simulation results with this
value proved satisfa tory.

The weighting matri es R, Q. The weighting matri es upon the outputs and
ontrol in rements are important design parameters. Both give a measure of
the tra king properties required from the losed loop system.
Sin e the referen es for the inner-loop have already been s aled, as part
of the inner loop design, su h that a unit hange of ea h referen e is equally
signi ant, it is possible to set R = diag (1; 1), and avoid tuning this weight
alltogether. This leads to onsiderable simpli ation of the tuning pro edure.
The tuning is an iterative pro ess typi ally starting with Q = diag(1; 1). The
rst step in the Q ontroller parameter design was to tune it in the absen e
of onstraints. This tuning of Q is not a one step pro ess. At rst we tune
using the time simulations involving the linear model of the plant. This step is
followed by ne tuning, done by time simulations employing the full nonlinear
model of the plant.
In order to improve the passenger omfort, whi h means that the ontrol
is less tight, we have to redu e Q. Be ause the H1 loop-shaping ontroller

389
redu es the amount of un ertainty in the inner-loop we do not require so mu h
robustness from the MBP C ontroller. Hen e this allows small values for Q.
The nal value of the output weighting matrix was Q = diag(0:007; 0:02).
On e we have de ided the initial tuning parameters we an pro eed to time
simulations. For the rst stages of the tuning pro edure it is re ommended to
have a short simulation time (six up to ten times the maximum time onstant
of the plant) and at the beginning to start in the un onstrained ase and then
to move towards the onstraint one.

Software tools for ontroller synthesis and analysis

Software whi h we all a Development Spa e was onstru ted using MATLAB
SIMULINK in order to enable us to perform analysis and simulations with var-
ious MBP C ongurations. This environment uses linear state-spa e ontin-
uous or dis rete MBP C internal models although the simulation plant model
an be nonlinear. It handles non square systems, disturban es and an in-
lude state-estimation of the plant model as an external fa ility by adding the
orresponding SIMULINK blo k. Expli it onstraints on manipulated input
hanges, input variables, outputs and states are handled. It allows the user
to handle un onstrained problems, hara terised by an analyti solution, and
onstrained problems solved by means of hill limbing algorithms.

Clock

Reference
Plot facilities
Double click block to plot
the variables involved in
MBPC control
Print graphs
Double click block MBPC Aircraft
to print the graph window
Print graphs to file The Plant Model
Double click block to save as Double click block
"results.ps" the graph window to edit the model parameters
Save results to file START UP
Double click block to save Double click block
the variables in "results.mat" to load initial parameters
MBPC Tuning Parameters Frequency plots
Double click block while Double click block to plot
the simulation is running the frequency plots in
to change the the unconstrained case
MBPC parameters

Figure 25.7: The SIMULINK MBPC Development Spa e

The environment shown in Figure 25.7 provides fun tions and has features
that give the user apability to design and simulate MBP C ontrollers for
simple and omplex multivariable plants. Only state spa e internal models

390
are used be ause of the reliability of the numeri al algorithms involved, and
the possibility of interfa ing them with other state-spa e based tools. su h
as the subspa e method  used in identi ation. The tuning parameters of
the MBP C method are passed immediately to the MBP C algorithm without
waiting for a simulation run to nish. This feature an be used to understand
the inuen e of the various parameters and the way of tuning them.
The main requirement of the Development Spa e is that the state mea-
surements from the plant model are available if estimation of them is not used.
When a redu ed model of the plant is employed, as the MBP C internal model,
a state transformation is needed in order to provide the redu ed model, em-
ployed for predi tion, with the ne essary initial measurement. Therefore, in the
SIMULINK plant blo k we have in luded the ne essary transformation matrix.

25.3.3 Analysis of the longitudinal hannel autopilot

At this stage a trial and error pro ess omplemented by the designer's knowl-
edge have been performed in order to a hieve the required performan e for the
ontrolled plant. The performan e was tested with the full nonlinear model by
providing steps in the referen e signal of the ommanded altitude and velo ity.

The altitude response for various aircraft configurations The crosscoupling in airspeed for various aircraft configurations
10 0.6

5
0.4
The airspeed and the airspeed reference
The altitude and the altitude reference

0
0.2

5
0

10

0.2
15

0.4
20

0.6
25

30 0.8
0 20 40 60 80 100 120 0 20 40 60 80 100 120
Time t [s] Time t [s]
The engine throttles actuators The tailplane actuators
0.095 0.07

0.08
0.09
0.09

0.085 0.1

0.11
rad

rad

0.08
0.12

0.075 0.13

0.14
0.07
0.15

0.065 0.16
0 20 40 60 80 100 120 0 20 40 60 80 100 120
Time t [s] Time t [s]

Figure 25.8: Results of the longitudinal analysis in the ase of MBP C as a

Flight Management System (altitude 30m step response)
In the altitude ase the ontrolled air raft was subje ted to a step of 30m for

391
various gravity entre positions, mass variations and time delays (as des ribed
in Se tion 14.3.3). The altitude response for various ombinations of these,
together with the ross ouplings in air speed and orresponding a tuator (the
engine throttle and tail-plane) movements are shown in Figure 25.8. The step
tra king is within the spe i ations. The biggest variation from the nominal
ase being 4m. The ross oupling between altitude and airspeed is within
bounds (eg. smaller than 1kt). The ontrol a tivity is limited within the
onstraints imposed for the MBP C design.
Conversely, he king the airspeed response subje t to a step of 13 m/s we
an on lude that the response is satisfa tory in the fa e of the same un er-
tainties as in the previous ase, see Figure 25.9.The traje tory following of the
nonlinear air raft was studied using time simulations for various ongurations
of the plant.

Channel Responses Performan e a hieved De ision

Longitudinal altitude tr = 30 s, ts = 50 s, Mp = 0% +,,+
velo ity tr = 4 s, ts = 30 s, Mp = 5% +,+,+

Table 25.4: Analysis results a hieved with the nonlinear longitudinal plant

As shown in Figures 25.8, 25.9 and Table 25.4 results were satisfa tory.
The use of the a priori information on the referen e traje tory an be observed
when the re eding horizon me hanism brings it into the predi tion horizon of
the MBP C ontroller.

In Figure 25.10 we depi t time responses of the RCAM longitudinal han-

nel full non-linear model. MBP C was implemented with onstraints pla ed
upon the a tuator dee tions and rates and on the outputs as ight envelope
restri tions.

We have subje ted the ontrolled air raft to dierent s enarios, relevant for
the longitudinal ase (Figure 25.10). At 20 s the RCAM goes into a des ent at
a rate of 10 m/s. At 10 s there is a wind-shear of 10 m/s for a 10 s duration.
The predi tion horizon of the MBP C is N2 = 10 s. The H1 ontroller by
itself tries to re over from the disturban e as fast as possible and return to
the original altitude of 1000 m. The ombined MBP C /H1 stru ture, though,
takes into a ount that the air raft is going to start des ending at 20 s and
does not try to rea h the original set-point hen e improving passenger omfort.
In ee t the MBP C ontroller is modifying the referen e to the inner losed
loop. At the bottom of the des ent there is no overshoot for similar reasons.

The transients in the rst 5 s (Figure 25.10) are due to the non-linear model
not being perfe tly trimmed. The ombined ontroller minimises the overshoot
and follows the referen e well within the design spe s from Se tion 14.3.2. At
75 s there is a head wind of 3 m/s that lasts for 17 s. As the predi tion horizon
is 10 s it an be dedu ed that the disturban e reje tion apability of both
ontrollers is the same in these ir umstan es.

392
 The airspeed response for various aircraft configurations The crosscoupling in altitude for various aircraft configurations
14 8

12
6

The airspeed and the airspeed reference

The altitude and the altitude reference

10

4
8

6 2

4
0

2
0

2 4
0 20 40 60 80 100 120 0 20 40 60 80 100 120
Time t [s] Time t [s]
The engine throttles actuators The tailplane actuators
0.18 0.04

0.16 0.06

0.14 0.08
rad

rad

0.12 0.1

0.1 0.12

0.08 0.14

0.06 0.16
0 20 40 60 80 100 120 0 20 40 60 80 100 120
Time t [s] Time t [s]

Figure 25.9: Results of the longitudinal analysis in the ase of MBP C as a

Flight Management System (airspeed 13 m/s step response)

Altitude descend response Comparative results Hinf (dashdot) and MBPC+Hinf (solid) controllers Speed response Comparative results Hinf (dashdot) and MBPC+Hinf controllers (solid)
1050 85

1000
80
950

900
75
850
speed [m/s]
altitude [m]

800 70

750
65
700

650
60
600

550 55
0 10 20 30 40 50 60 70 80 0 50 100 150
time [s] time [s]

Figure 25.10: Altitude and airspeed response (subje t to wind-shear) ompar-

ative results: dash dot the H1 ontroller and solid the MBP C /H1 ontroller

393
25.3.4 Results of the automated evaluation pro edure
The RCAM design hallenge involves designing a ontrol law that is able to
perform an approa h to landing in the presen e of turbulen e and wind-shear,
whilst remaining robust to mismodelling 14.3.
This Se tion presents the methodology-independent results of the designed
ontroller. It is mostly based on the evaluation mission and s enario dened in
Se tion 14.3.3. Both `overall tra king performan e' and `inner-loop behaviour'
of the ontrolled system will be evaluated by means of bounds on key variables.
A dis ussion of the behaviour of the ontrolled air raft will be based on the
relevant ight segments for the longitudinal hannel. A omparison between
the guidan e and stabilisation fun tions and the overall autopilot is provided.
Finally, a summary of the omparative numeri al results of the evaluation will
be presented.
To prove the idea of the ombined ontroller stru ture the most relevant
segment of the approa h manoeuvre is extra ted. This segment, that represents
the nal des ent to land, requires the apture of 6 and 3 glide-slopes.
We start with a glide-slope of

6 ; again it is unavoidable that the air raft
leaves the desired traje tory. It should return to the traje tory without over-
shoot and well within a period of 30 s. After that we go to a glide-slope 3 of
su h that we get an inverse behaviour with respe t to the desired traje tory,
that should be about half the size of the rst response (if the system has a
more or less linear behaviour). In Figure 25.11 the longitudinal response of the
air raft is plotted for three entre of gravity lo ations and with bounds that
spe ify a eptable behaviour.
Both ontrollers, MBP C /H1 (dotted) and H1 on its own (dash-dot),
behave in an a eptable manner. The dieren e arises when onsidering the
verti al deviation from the desired glide slope. While the H1 ontroller tries
to follow the des ent referen e traje tory as losely as possible resulting in
overshoots, the MBP C /H1 ontroller takes advantage of the a priori known
traje tory, optimising and improving omfort and safety. The verti al devia-
tions from the desired glide-slope are plotted in Figure 25.11.
While on nal approa h with a glide-slope of 3 the ee t of a wind-shear
is onsidered (see segment IV gh in Figure 25.11, se ond graph). The verti al
deviations from the desired glide-slope are plotted.
Table 25.5 gives omparative numeri al results based on the two simulations
with the two distin t ontrollers, the ombined autopilot MBP C /H1 and
the H1 loop shaping ontroller on its own. Ea h segment of the evaluation
pro edure is onsidered via the design riteria: omfort, safety and power.
The performan e and robustness riteria addressed by the inner ontroller are
not in luded. Be ause the evaluation riteria are independent of the type of
ontroller used the table ontains al ulable indi ators that enable us to obtain
an obje tive omparison between this one and ompletely dierent ontrollers
from other design hapters.
For ea h of the traje tory segments a single number was al ulated. The
smaller the numbers the better the design. The motivation and al ulation

394
 Third segment: altitude deviations Fourth segment: altitude deviations
30 30

20 20

altitude deviation [m]

altitude deviation [m]

3
10 2 f 3 10
4
g h
0 0

10 e 10

20 20

30 30
16 15 14 13 12 11 11 10 9 8 7 6 5 4 3 2 1
xposition (XE) [km] xposition (XE) [km]
Third and forth segment: The engine throttle actuators Third and forth segment: The tailplane actuator movement
0.08
0.14

0.12

0.1 0.1

0.08
rad

rad

0.06
0.12
0.04

0.02

0 0.14
300 350 400 450 500 550 300 350 400 450 500 550
Time t [s] Time t [s]

Figure 25.11: Segment III: verti al deviations from the desired glide-slope and
Segment IV: verti al deviations from the desired glide-slope and orresponding
a tuators movements (The MBPC/H-inf ombined autopilot (dotted) and the
H-inf ontroller (dash-dotted)

Criteria MBP C /H1 H1

Comfort 0.5137 1.3316
Safety 0.0075 0.0084
Power 0.0155 0.0150

Table 25.5: Comparative numeri al results

395
prin iples of these gures an be found in Se tion 14.3.3.

25.4 Con lusions and Lessons Learned

The main di ulty in designing MBP C is in the tuning of the many param-
eters available in the algorithm. But we found that, with experien e, partly
systemati tuning pro edures were developed, also see [151.
It was possible to obtain a satisfa tory ontroller despite a very simple
model used as an internal model by the MBP C . In order to move towards
more systemati design, from the robustness point of view, it was ne essary to
use an inner robust stabilising H1 loop-shaping ontroller to ensure robustness
against large perturbations (eg. delays, mass and gravity entre variations).
The absen e of well dened rules for hoosing the tuning parameters requires
serious experien e of the designer in order to rea h a eptable results. But
assuming a good analysis tool set and adequate rules of thumb, the problems
an be over ome. It may be noted that this problem is not spe i for MBP C ;
for example the learning urve that will bring a lassi al ontrol engineer to the
stage when a omplex MIMO design an be su esfully designed using H1 will
take up to several months, as dis ussed in Chapter 7. The dieren e is that
for the H1 loop-shaping ontroller synthesis more systemati pro edures have
been developed. On e the above mentioned obsta les are over ome, a fairly
systemati redesign pro ess an be developed, if the designer who inherits the
redesign has experien e of the te hnique and of the pro edures used at the rst
stage of design.
The onstrained optimisation, whi h has to be solved on-line, augments
the omputational omplexity of implementing the ontroller. The di ulty to
be over ome was redu ing the omputation time su iently to allow real-time
operation. This was ensured by the low sampling rate employed for the outer
MBP C ontroller. Moreover, progress an be expe ted both in the e ien y
of solution algorithms for MBP C and in the power of the hardware on whi h
they run.
Constrained MBP C may not oer any advantages over more onventional
ontrol algorithms apart from straightforward onstraint handling. As a result
we have presented in Se tion 25.3 an automati pilot based on a ombined
use of an H1 loop-shaping ontroller that will provide the stability augmenta-
tion and guidan e fun tions and an MBP C ontroller that will a t as a ight
manager and overall supervisor. As expe ted, the MBP C proved to be an
ee tive te hnique for ight ontrol, on e this higher-level obje tive of ight
management was in luded.
Current quali ation and erti ation pro edures are not appropriate for
MBP C or some other modern ontrol solutions. As long as onstraints are
not a tive our solution, having an analyti al form, an be ertied just as well
as any other linear ontrol law. But that provides only a partial analysis of the
ontroller, when it is operating in its linear mode. Moreover, it ould be argued
that if the autopilot goes unstable the MBP C ontroller ould be swit hed o.

396
A knowledgements
The work referred in Se tion 25.3 was developed in ollaboration with our
olleague George Papageorgiou [192. We are grateful for the omments given
by the evaluators regarding our Report [119: Prof. Rudolf Bro khaus (DASA),
Dr. M.P.S hifaudo (ALENIA), Dr. J.F.T.Bos (NLR), Dr. G.S hram (DUT),
Dr. R.de Vries (DUT). Mihai Huzmezan is supported by Pembroke College
Cambridge, the Lundgren fund, ORS s holarship, Cambridge Overseas Trust
and CT Taylor fund.

397
26. A Fuzzy Control Approa h

Gerard S hram and Henk B. Verbruggen1

1

Abstra t. Pilot heuristi s of ying an air raft are implemented in

the design of a fuzzy logi ontroller (FLC). The FLC design ex-
ists of longitudinal and lateral outer loop tra king ontrollers om-
bined with lassi al inner loop attitude ontrollers. The rst step
in ludes tuning the inner loop ontrollers using lassi al ontrol
te hniques. Next, the FLCs outer loop ontrollers are initialized
as linear, MIMO ontrollers in whi h the rules are derived from
ontrol strategies based on the experien e of pilots. The nal step
in ludes ne-tuning of the FLCs by modifying the rules based on
an initial evaluation of time responses with respe t to the spe i-
 ations. Additional rules for low airspeed and engine failure are
in luded as well, whi h show that gain s heduling and ex eption
handling an be in orporated in a straightforward way. Following
this three-step pro edure, a ompromise is found in whi h perfor-
man e and robustness properties are met at the ost of a slightly
too high verti al and lateral a eleration. The obtained FLC is of
low order, deterministi and transparent.

26.1 Introdu tion

Re ently, mu h attention has been paid to the appli ation of knowledge-based
ontrol te hniques for ight ontrol [228, 230. It shows that te hniques like
neural networks and fuzzy systems an provide appropriate tools for nonlin-
ear identi ation [156, 204, ontrol of high performan e air raft [183, 229,
heli opters [195, 233, and spa e raft [26, 106, ight ontrol re onguration
[142, 182, 263, and advisory systems [111, 232. In these appli ations, neural
networks generally serve as nonlinear, sometimes adaptive, fun tion approxi-
mators, while fuzzy systems are used as supervisory, expert systems.
The RCAM problem as formulated in Chapter 14 is a unique hallenge to
investigate the feasibility of knowledge-based ontrol te hniques for a realisti
ivil ight ontrol problem. Be ause the problem is relatively well-understood,
pilot heuristi s are available and linearized air raft models an be derived.
Therefore, a hybrid ontroller stru ture is proposed. The inner loop ontrollers
onsist of lassi al attitude ontrollers whi h are tuned using the air raft models
(see also Chapter 15). Then, for outer loop (tra king) ontrol, longitudinal and

1
Department of Ele tri al Engineering, Delft University of Te hnology. P.O.Box 5031,
2600 GA Delft, The Netherlands. Email: {g.s hram}{h.b.verbruggen}et.tudelft.nl

398
lateral Fuzzy Logi Controllers (FLCs) are developed. For this purpose, the
pilot heuristi s of ying an air raft are used. Experien ed human pilots know
how to handle an air raft satisfa tory, i.e. they know how to ombine om-
peting riteria su h as performan e, robustness, safety, and passenger omfort
without mathemati al formulations. Fuzzy logi provides a transparent inter-
fa e between the low-level, attitude ontrol of an air raft and the high-level
reasoning of human pilots. The knowledge is aptured through the use of if-
then rules and linguisti terms like small, big, et . for ea h variable. Additional
rules for low airspeed and engine failure are in luded as well, whi h shows that
gain s heduling and ex eption handling an be in orporated straightforwardly
in the same framework.
A tutorial on FLC is provided in Chapter 13. This hapter introdu es the
ontroller stru ture in se tion 26.2, followed by the dis ussion in se tion 26.3 on
how to translate the requirements whi h are imposed on the RCAM problem
into the ontroller design. Next, the ontroller is tuned in se tion 26.4. An
extensive evaluation is given in se tion 26.5, and in se tion 26.6 the landing
results are shown. Finally, on lusions follow in se tion 26.7. A detailed report
on the FLC design an be found in [207.

26.2 Sele tion of the Controller Ar hite ture

26.2.1 Classi al inner loop ontrollers
The design starts with a pit h attitude ontroller. In the lassi al ontroller,
both pit h angle and pit h rate are fed ba k to the ontrol input of the tailplane
by the stati gains Kq and K , respe tively. An estimate of the pit h angle is
obtained by integrating _(t) = q(t) os((t)) r(t) sin((t)). In Figure 26.1 the
ontrol stru ture is shown.


- - K - -
+ + T air raft

-
6 -
6 -

 q
Kq

Figure 26.1: Blo k-s hemati representation of pit h attitude ontrol system.

Then, two lateral inner loops are sele ted: a roll attitude ontrol system,
and a yaw damper. In Figure 26.2 the roll attitude ontrol system is shown:
proportional feedba k to aileron in luding roll rate damping by the gains K
and Kp , respe tively. In Figure 26.3, the yaw damper is shown. The fun tion of
the yaw damper is to in rease the damping of the Dut h roll motion. In the yaw
rate feedba k loop, a washout lter is added in order to allow a onstant yaw
rate in ase of a o-ordinated turn (while sideslip is redu ed to zero, see outer
loop ontroller). An extra (washed out) feedba k, whi h is in luded like in the
lassi al ontrol approa h ( hapter 15), is the gain KR between roll angle 
and rudder dee tion R . This feedba k de reases the lateral a eleration in

399
 ase of roll attitude hanges.


- - K - -
+ + A air raft

-
6 -
6 -

 p
Kp

Figure 26.2: Blo k-s hemati representation of roll attitude ontrol system.
s
s+1 KR 

R -? -
+ - R air raft
r
6 s
-

s+1 Kr 
Figure 26.3: Blo k-s hemati representation of yaw damper.

26.2.2 Fuzzy logi outer loop ontrollers

The attitude ontrollers serve as inner loops for the longitudinal and lateral
tra king ontrollers. Suppose that the longitudinal task of the pilot is to ontrol
the altitude and velo ity of the air raft. The ontrol behaviour an be related
to the total energy of the air raft [147. For example, if the pilot realizes that
the air raft ies too low and too slow, he or she will in rease thrust in order to
in rease the energy of the air raft. On the other hand, if the air raft ies too
low but the velo ity is too high, than the pilot will in rease the pit h angle.
Thrust in fa t in reases the total energy, while an ex hange between kineti
and potential energy an be a hieved by pit h angle hanges (via the tailplane).
Based on these heuristi s, two rule bases are dened for throttle setting
and pit h angle ommands, respe tively (Tables 26.1 and 26.2). The labels
are dened negative (N), positive (P), and extreme (E), very big (VB), big
(B), medium (M), small (S), very small (VS), and zero (ZE). The number of
rules is hosen quite large in order to make lo al modi ations of the ontroller
possible. The terms are related to a series of membership fun tions between
[ 1; 1. The position and shape of the membership fun tions are dened su h
that the initial input-output fun tion of the FLC is linear between the bounds
(see Chapter 13). The membership fun tions are shown in Figure 26.4.
In Figure 26.5, the longitudinal ontroller is shown. The air raft blo k
represents the air raft dynami s in luding the inner loop pit h angle ontroller.
The fa tors S used for s aling big and small errors are translated into the
[ 1; 1 domain, and ontrol ommands [ 1; 1 are s aled towards minimum
and maximum ontrol a tions. The ontrol a tions are throttle settings T H (=
T H1 = T H2 ) and pit h angle ommands  . It is well-known that human pilots
are able to mimi derivative and integral a tions as well [167. The derivative
signals of altitude and velo ity are required for extra damping. For this purpose,

400
Table 26.1: Rule base for throttle setting, e.g. if velo ity error is PM (air raft
too slow), and altitude error is ZE, then throttle hange is PS (more thrust).

velo ity error (!)

altitude error (#) NB NM NS ZE PS PM PB

NB NE NVB NB NM NS NVS ZE
NM NVB NB NM NS NVS ZE PVS
NS NB NM NS NVS ZE PVS PS
ZE NM NS NVS ZE PVS PS PM
PS NS NVS ZE PVS PS PM PB
PM NVS ZE PVS PS PM PB PVB
PB ZE PVS PS PM PB PVB PE

Table 26.2: Rule base for pit h angle ommands, e.g. if velo ity error is PM,
and altitude error is ZE, then pit h angle hange is NS (nose little down).

velo ity error (!)

altitude error (#) NB NM NS ZE PS PM PB

NB ZE NVS NS NM NB NVB NE
NM PVS ZE NVS NS NM NB NVB
NS PS PVS ZE NVS NS NM NB
ZE PM PS PVS ZE NVS NS NM
PS PB PM PS PVS ZE NVS NS
PM PVB PB PM PS PVS ZE NVS
PB PE PVB PB PM PS PVS ZE

limb rate and horizontal a eleration are used. The integral signal paths are
required for zero steady-state errors. The integral and derivative gains are
simply set at ommonly used values in lassi al ight ontrollers: Kd = 5 and
Ki = 0:1. Extra tuning of these values did not appear to be ne essary. Noti e
that the s aling, and the derivative and integral ompensation are performed
in the rst and last blo k of Figure 13.3.
Pilot heuristi s are again onsidered for the lateral outer loop ontroller.
The sele ted variables to be ontrolled are heading angle , lateral deviation
ylat , the sideslip angle  and heading rate _ for o-ordinated turns. A human
pilot will generally handle heading and lateral deviation by roll/aileron om-
mands, and sideslip by rudder dee tion. This strategy is implemented in two
FLCs. For a desired heading rate _ in ase of a o-ordinated turn, an extra
roll of sin 1 ( Vg
_ ref ) is added separately.
Lateral deviations must rst be translated into roll angle ommands. Sup-
pose that a lateral deviation ylat is present. In order to redu e the lateral
deviation, we want to initiate a negative lateral velo ity y_ of e.g. a fa tor
10 smaller than lateral deviation. Then the lateral deviation will vanish in
about 30s. Be ause sin  Vy_ (for small heading angles  ), an extra desired
heading angle hange an be determined by
= 101
V1
ylat .
In Table 26.3, a rule base is dened with heading error e = ref and
the extra desired heading hange
as ante edents, and roll angle ommand

401
 input: altitude error input: airspeed error

NB NM NS ZE PS PM PB NB NM NS ZE PS PM PB
1 1

0.5 0.5

0 0
1 0.5 0 0.5 1 1 0.5 0 0.5 1

outputs: pitch angle and throttle

NE NVB NB NM NS NVS ZE PVS PS PM PB PVB PE

1

0.5

0
1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1

1 + Kd s 
Figure 26.4: Membership fun tions for longitudinal FLC.

- ?- SV -
VA;ref + - [-1,1
- S - 1 + Ks
[-1,1
- i

+ -VA
- - Sh - -S - 1 + Ks T H- -h
air raft
href + FLC
i inner loops

6 -
TH

1 + Kd s 
Figure 26.5: Blo k-s hemati representation of longitudinal ontroller.

as onsequent. The rule base for sideslip error e = ref  is shown as well.
Table 26.3: Rule bases for roll angle and rudder ommands.

e (!) sideslip rudder

(#) NB NS ZE PS PB
error e R
NB NE NVB NB NS ZE NB NM
NS NVB NB NS ZE PS NS NS
ZE NB NS ZE PS PB ZE ZE
PS NS ZE PS PB PVB PS PS
PB ZE PS PB PVB PE PB PM

The membership fun tions for heading and roll angle are shown in Fig-
ure 26.6. With the fuzzy sets whi h are used in the rule base, an initial linear
fun tion is obtained in the interval [ 1; 1. The other indi ated membership
fun tions will be used in a later stage. For the sideslip error and rudder de-
e tion, the membership fun tions are shown in Figure 26.7. The fuzzy sets
(NM,NS,ZE,PS,PM) are on entrated between [ 0:25; 0:25. In se tion 26.5,
it will be shown that in ase of an engine failure, the membership fun tions
(NVB,NB,PB,PVB) must be used to represent the more aggressive ontrol be-
haviour of the (human) pilot.
In Figure 26.8, the lateral ontroller is shown. The S -gains are s aling
parameters as part of fuzzi ation and defuzzi ation. Noti e that an extra
integrator ( Ki = 0:2) is added for zero steady-state sideslip suppression.

402
 input: heading error input: delta heading

NB NS ZE PS PB NB NS ZE PS PB
1 1

0.5 0.5

0 0
1 0.5 0 0.5 1 1 0.5 0 0.5 1

output: roll angle

NE NVB NB NM NS NVS ZE PVS PS PM PB PVB PE

1

0.5

0
1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1

Figure 26.6: Membership fun tions for lateral tra king ontroller.
output: rudder deflection
input: sideslip

NVB NB NM NS ZE PS PM PB PVB
NB NS ZE PS PB 1
1

0.5 0.5

0 0
1 0.5 0 0.5 1 1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1

Figure 26.7: Membership fun tions for sideslip ontroller.

26.3 Translation of Criteria into Design

In the following, guidelines are provided on how to translate the design riteria
as dened in se tion 14.3 into the ontroller design (see also se tion 13.4).

 The desired performan e is addressed in two ways: the hoi e of the

ontroller stru ture and the tuning pro edure. A natural bandwidth sep-
aration is a hieved by fast inner loops and slow outer loops. This
means that the desired response time of roll and pit h angle ommands
are dened a fa tor 4 to 5 smaller than for tra king. Furthermore, (small)
integrators are added to ensure that altitude and airspeed steady-state
errors are zero. Lateral deviation a ts as an integral signal for heading.
Next, the ontroller is tuned based on time responses under nominal on-
ditions. Tuning the s aling fa tors of the inputs of the FLCs determines
the overall gain of the losed loops. For example, if the time responses
are in general too slow, the s aling fa tors are in reased. Individual rules
are adjusted su h that lo al desired hanges are a hieved, e.g. more ag-
gressive ommands around a set-point. Finally, by adding new rules for
ertain onditions, impli it gain s heduling an be a hieved, e.g. in ase
of an engine failure.

 Robustness with respe t to enter-of-gravity and mass variations appeared

to be no problem for the FLCs. Generally, sin e human operators often
perform as low gain ontrollers, the obtained FLCs are quite robust.

403
 _ ref
- sin 1 ( Vg
)
ylat - 1 - y_ 1
-
10 V S - [-1,1

- -S - - S - ? - -
FLC
ref + [-1,1 + + 
6- [-1,1
air raft
+ inner
ref - - S -
+ [-1,1
- SR - 1 + Ks R-i loops
-
 6
FLC [-1,1
-

Figure 26.8: Blo k-s hemati representation of lateral ontroller.

Furthermore, the pilot is able to adapt his or her ontrol behaviour; a

strategy whi h an be implemented in the FLC. In the heading FLC for
instan e, it appeared to be ne essary to dene separate rule bases for low
airspeed (impli it, smooth gain s heduling).

 The ride quality requirements are essentially a eleration limits and suf-
 ient damping. The verti al and lateral a eleration are losely related
to pit h rate and roll rate, respe tively. In order to a hieve low lateral
a eleration, a ross-link from roll angle to rudder dee tions is added.
In order to de rease the verti al a eleration, a pit h rate limiter ould
be added. In Chapter 15, these problems are dis ussed in detail.
In order to a hieve damped responses, derivative signals in the form of
horizontal a eleration and limb rate are added in the longitudinal outer
loop ontroller. Unfortunately, lateral a eleration is not assumed to be
known, and annot be used for damping of lateral responses like heading.

 Con erning the safety requirements, no spe ial attention is paid to stall
speed and angle of atta k prote tion. If ne essary, an angle of atta k
prote tion an be added in the pit h inner loop ontroller, and a stall
speed prote tion in the related FLC. Furthermore, the roll angle limit
is simply a hieved by s aling the FLC su h that {-1,1} orresponds to
the minimum and maximum values. Finally, sideslip angle redu tion is
addressed by adding an integrator in the sideslip angle outer loop.

 Finally, ontrol a tivity appeared to be a minor problem. As mentioned

earlier, the ontroller is mainly of low gain. Additional tuning of ontrol
a tivity an be addressed dire tly by the rule bases of throttle a tivity
and rudder dee tion, and indire tly by the rule bases for pit h and roll
angle, e.g. less aggressive ommands.

It is lear that the translation of human heuristi s into a FLC is a systemati

pro edure. This fa t is also re ognized by industry, and re ently eorts have

404
in reased to dene a European industry standard for the development method-
ology of fuzzy logi systems, based on the ISO-9000 general system develop-
ment guidelines [248. However, ne-tuning the performan e of the ontroller
is a matter of trial-and-error like in lassi al ontrol, but using the provided
guidelines and an understanding of the inuen e of ontroller parameters (see
Chapter 13), a satisfa tory ontroller an be obtained.

26.4 Tuning of the Controller Parameters

The hybrid ontroller, is tuned in three steps. First the inner loop ontrollers
are tuned. Then the s aling fa tors of the initial FLCs are determined. Finally,
the FLCs are improved by ne-tuning the rule onsequents and eventually
adding new rules.

Step 1: Tuning the lassi al ontrollers

In Chapter 15, it is shown how root lo us te hniques for tuning of lassi al
K =
inner loop gains an be used. For the pit h attitude ontroller, values of
1 and Kq = 1:5 are found for appropriate damping of phugoid motion and
a response time of 5s. Good yaw rate damping is a hieved with gain values
Kr = 2 and a washout lter time onstant  = 2s. A roll angle response time
of 5s is a hieved with the gains K = 1:5 and Kp = 1:5. Finally, lateral
a eleration in ase of roll angle ommands is minimized with the gain from
roll to rudder KR = 0:4, and a washout lter time onstant  = 2s.

Step 2: Tuning the initial FLCs

Tuning the initial FLCs of se tion 26.2 involves tuning the s aling fa tors. The
s aling fa tors of the FLC outputs T H S ;S S ;S
 and R  are dened su h that
minimum and maximum allowed values orrespond to the {-1,1} limits. For
throttle setting: 0.5 to 10 degrees, for pit h angle -10 to 20 degrees, and for roll
angle and rudder dee tion -30 to 30 degrees. In the implementation phase,
the integrators are pla ed in front of the s aling fa tors.
Next, the s aling fa tors of the FLC inputs must be dened. A rst hoi e
Sh 1
for the s aling fa tor of altitude is
100 . In other words: an altitude error
of 100m is assumed to be very big. An airspeed hange of 12 m/s results in
approximately an equal total energy hange
2 as an altitude hange of 100m.
Therefore, a fa tor of SV =
1 VA = 80 m/s).
12 is dened (nominal airspeed
The initial longitudinal FLC is tested with a simultaneous step on altitude
and airspeed. The results show that the response time is too large. Therefore,
the s aling fa tors are in reased to Sh = 251 and SV = 31 . Noti e that the
s aling fa tors determine the overall gain of the altitude and airspeed loop.
The ee t is plotted in Figure 26.9.

2
Total energy is dened as mgh + 1=2mV 2 . For onstant total energy, the following
equation an be derived: g

h = V

V + 1=2(
V )2 .

405
 Airspeed response Altitude response
81 1030

80 Sh=1/100,Sv=1/12
1020
79 . Sh=1/25,Sv=1/3
1010
78
1000
77

76 990
0 20 40 60 0 20 40 60

Figure 26.9: Ee t of tuning the s aling fa tors Sh and SV .

Similarly, the s aling fa tors for heading and sideslip are determined: S =
180 1 S = 180 1
 5 and
 5. Errors of 5 degrees are here interpreted as big errors.

Step 3: Fine-tuning the FLCs

The initial longitudinal ontroller appeared to show good performan e ex ept
(1) too mu h overshoot for altitude responses, and (2) too mu h altitude de-
viation for a large airspeed hange. In order to improve the responses, the
following lo al modi ations are performed: (1) more throttle a tion around
set-point for faster settling, and (2) greater priority for altitude with respe t
to airspeed. The desired modi ations are a hieved by only hanging the rule
onsequents. The rule bases for throttle setting and pit h angle are shown in
Table 26.4 and 26.5 where the arrows indi ate the modied onsequents.
The ee t of the modi ations is shown in Figure 26.10. The ross- oupling
on altitude is de reased without deteriorating the airspeed step response. The

Airspeed response Altitude response

85 1010

80 initial FLC 1005

. tuned FLC
75 1000

70 995

65 990
0 20 40 60 0 20 40 60

Figure 26.10: Ee t of ne-tuning the rules.

ee t of ne-tuning the rules an be further analyzed by al ulating the output

of the FLC as a fun tion of altitude and airspeed errors. In Figure 26.11, the
outputs are shown: the rule modi ations introdu e lo al nonlinearities into
the initial, linear loop gain.

The initial lateral ontroller shows good responses and needs no further on-
sequent modi ations, ex ept in ase of an engine failure. The lateral deviation
of the ight path is too large and the roll angle ommand is larger than the
allowed 10 degrees. In general, in ase of an engine failure, the human pilot

406
Table 26.4: Modied rule base for throttle setting: the throttle ommands are
more aggressive around the set-point.

Ve (!)
he (#)
NB NM NS ZE PS PM PB

NB NE NVB NB NM NS NVS ZE
NM NVB NB NM NS NVS ZE PVS
NS NB NM !NM !NS ZE PVS PS
ZE NM NS !NS ZE !PS PS PM
PS NS NVS ZE !PS !PM PM PB
PM NVS ZE PVS PS PM PB PVB
PB ZE PVS PS PM PB PVB PE

Table 26.5: Modied rule base for pit h angle: the onsequents hanges are
dened su h that ontroller a tions with respe t to altitude errors are more
aggressive, and with respe t to velo ity errors less aggressive.

Ve (!)
he (#)
NB NM NS ZE PS PM PB

NB !NVS !NS !NM !NB !NVB !NE NE

NM !ZE !NVS !NS !NM !NB NB NVB
NS !PVS !ZE ZE NVS NS !NS !NM
ZE !PS !PVS PVS ZE NVS !NVS !NS
PS !PM !PS PS PVS ZE !ZE !NVS
PM PVB PB !PB !PM !PS !PVS !ZE
PB PE !PE !PVB !PB !PM !PS !PVS

will in rease the power of the remaining engine in order to keep altitude and
airspeed (energy on ept), and he will add extra rudder dee tion in order to
keep the desired heading angle with minimal sideslip. Rudder dee tion in ase
of motor failure is a general pilot pro edure. The power in rease is a hieved by
the longitudinal FLC based on energy prin iples. The extra rudder dee tion
is in prin iple also a hieved with the lateral ontroller. However, the human
pilot will rea t more aggressively than for the ase of no engine failure, with
respe t to heading and sideslip errors. The pilot behaviour is modelled by two
extra rule bases, depi ted in Table 26.6. The onsequents are more aggressive
than in the rule bases for no engine failure (Table 26.3). The ee t of the
modi ations is that the gains of the lateral outer loops are in reased.

Table 26.6: Rule bases in ase of engine failure.

e (!) sideslip rudder

(#) NB NS ZE PS PB
error e R
NB NE NE NVB NM ZE NB NVB
NS NE NVB NM ZE PM NS NB
ZE NVB NM ZE PM PVB ZE ZE
PS NM ZE PM PVB PE PS PB
PB ZE PM PVB PE PE PB PVB

407
 throttle command
pitch command

12 30

10
20
8
10
6
0
4

2 10

0 20
40 40
20 5 20 5
0 0
0 0
20 20

altitude error (m) 40 5 altitude error (m) 40 5

airspeed error (m/s) airspeed error (m/s)

Figure 26.11: FLC outputs as fun tion of airspeed and altitude errors.

Smooth swit hing between the rule bases for engine ON and OFF is auto-
mati ally a hieved by dening two membership fun tions for the extra input
engine failure. The membership fun tions are shown in Figure 26.12. In ase
of an engine failure, the rules for engine OFF are dire tly initiated by swit hing
from 0 to 1. When the engine is restarted, the pilot swit hes gradually its on-
trol behaviour from engine OFF to engine ON sin e he/she has to re over rst.
In Figure 26.13, the extra input signal for the ontroller is shown: the engine
breaks down at t=0s and is restarted at t=100s. The ee t of the extra rules
for engine failure is also shown. The lateral deviation and roll angle responses
are de reased by the more aggressive rudder ommands.
input: engine failure input: airspeed

ON OFF LOW NOMINAL

1 1

0.5 0.5

0 0
0 0.5 1 60 65 70 75 80

Figure 26.12: Membership fun tions for engine and airspeed inputs.

engine failure signal lateral deviation roll angle

1.5 20 15

0 10
1
20 5

40 0
0.5
initial FLC
60 5
with extra rule base
0 80 10
0 50 100 150 200 0 50 100 150 200 0 50 100 150 200

Figure 26.13: Ee t of extra rule bases in ase of engine failure.

After tuning the ontroller for nominal onditions, robustness is studied.

408
The only real problem appeared to be airspeed variations for the lateral on-
troller. Using similar roll ommands for low and high airspeed indu es high
and low heading rates, respe tively, sin e _  g
Va
. In order to make the
lateral heading ontroller more robust against airspeed variations, extra rule
bases with smaller roll ommands are dened for low airspeed (Table 26.7).

Table 26.7: Rule bases roll angle for low airspeed.

e (!)

(#) NB NS ZE PS PB

NB NB NM NS NVS ZE
NS NM NS NVS ZE PVS (Engine ON)
ZE NS NVS ZE PVS PS
PS NVS ZE PVS PS PM
PB ZE PVS PS PM PB

e (!)

(#) NB NS ZE PS PB

NB NB NB NM NS ZE
NS NB NM NS ZE PS (Engine OFF)
ZE NM NS ZE PS PM
PS NS ZE PS PM PB
PB ZE PS PM PB PB

Smooth s heduling between the rule bases for low (60m/s, Vstall for m =
150.000kg) and nominal airspeed is a hieved by dening the two membership
fun tions whi h are depi ted in Figure 26.12. The ee t of the extra rule bases
is shown in Figure 26.14.

heading initial FLC heading with extra rules

8 8

6 6

4 4

2 2
nominal airspeed
0 low airspeed 0

2 2
0 20 40 60 0 20 40 60

Figure 26.14: Ee t of extra rule base for low airspeed on heading response.

26.5 Analysis of the Resulting Controller

The performan e and robustness of the ontroller is tested under variations
of the enter of gravity, mass, airspeed, and time delay. In Table 26.8, nine
onditions under whi h the ontroller is tested are dened. In the following
se tions, the longitudinal and lateral results are dis ussed.

409
Table 26.8: Flight onditions: variation of airspeed, enter of gravity ( .o.g.),
mass, and time delay.

No. Mass(kg) Airspeed(m/s) CGx ( ) CGz ( ) Time delay(s) Variation

1 120.000 80 0.23 0 0 nominal

2 120.000 80 0.15 0 0.1 .o.g.
3 120.000 80 0.31 0 0.1 .o.g.
4 120.000 80 0.23 0.21 0.1 .o.g.
5 100.000 80 0.23 0 0.1 mass
6 150.000 80 0.23 0 0.1 mass
7 120.000 67.4 0.23 0 0.1 1.3
Vstall
8 120.000 90 0.23 0 0.1 max. Va
9 150.000 80 0.31 0.21 0.1 worst ase

26.5.1 Analysis of longitudinal ontroller

First, in order to show that the ontroller indeed mimi s the pilot's behaviour
a ording to energy prin iples, an airspeed ommand of -3 m/s and an alti-
tude ommand of +25 m are simultaneously given under nominal onditions.
In Figure 26.15, the responses are shown. The a tuator responses show that
throttle setting remains pra ti ally onstant, and that by a tailplane dee tion
the energy is ex hanged from airspeed to altitude.

Airspeed response () Altitude response ()

81 1030 tailplane () and throttle ()
0.1
80
1020
0
79
1010 0.1
78
1000 0.2
77

76 990 0.3
0 20 40 60 0 20 40 60 0 20 40 60

Figure 26.15: Simultaneous airspeed and altitude ommand.

The performan e for only altitude hanges is tested by a step ommand of

-5m. The (nearly identi al) step responses are plotted in Figure 26.16. The
requirements are fullled for the 9 ongurations. Noti e the dieren es in
trimmed throttle and tailplane settings.

Altitude response tailplane () and throttle ()

1002 0.1

1000
0

998

0.1
996

994 0.2
0 20 40 60 0 20 40 60

Figure 26.16: Time responses altitude hange -5m.

410
 The ross- oupling ee t on velo ity is tested by a step response of -30m.
The maximum allowed velo ity deviation is 0.5m/s. The responses are plotted
in Figure 26.17. From the gure it an be on luded that the ross- oupling
spe i ation is met.

Altitude response Airspeed response

1010 0.4

1000
0.2
990
0
980
0.2
970

960 0.4
0 20 40 60 0 20 40 60

Figure 26.17: Time responses altitude hange -30m.

Next, the ontroller is tested for airspeed ommands. In Figure 26.18, the
responses are shown for an airspeed hange of -5m/s. Again, the spe i ations
on erning rise time, settling time and overshoot are met. Noti e the throttle
setting: the energy is slightly de reased orresponding to the lower airspeed.

Airspeed response tailplane () and throttle ()

2 0.1

0 0

2 0.1

4 0.2

6 0.3
0 20 40 60 0 20 40 60

Figure 26.18: Time responses airspeed hange -5m/s.

The ross- oupling ee t on altitude is tested by a step response of -13

m/s. The maximum allowed altitude deviation is 10m. Moreover, the airspeed
deviation must be smaller than 2.6 m/s after 15s. The responses are plotted in
Figure 26.19. Both requirements are met under all onditions.

Airspeed response Altitude response

5 1010

0 1005

5 1000

10 995

15 990
0 20 40 60 0 20 40 60

Figure 26.19: Time responses airspeed hange -13m/s.

Finally, in Figure 26.20, step responses for = 3 degrees are shown. Only

411
the settling time spe i ation is not met (30s instead of 20s). Be ause of the
integrator, the verti al deviation from the ight path is zero. Noti e that the
ight paths are dierent for low, nominal, and high airspeed.

Gamma response Altitude response

1 1100

0
1000
1
900
2
800
3

4 700
0 20 40 60 0 20 40 60

Figure 26.20: Time responses ight path hange -3 degrees/s.

The ride quality riteria are dened as maximum a elerations and min-
imum damping. Evaluating the verti al a eleration during large step om-
mands, indi ates that the maximum value of 0:05g is ex eeded with peak
values of 0:2g. The problem is that verti al a eleration is mainly aused by
pit h angle hanges. A rate limit on pit h angle ommands is a possible inner
loop ontroller solution (see Chapter 15). On the other hand, the damping of
the responses is very good. This was a hieved through the ompensation of alti-
tude and airspeed errors by limb rate and horizontal a eleration, respe tively.
Noti e therefore that the horizontal a eleration is very low as well.
During all simulations, the airspeed did not ex eed the 1:05
Vstall bound,
and the angle of atta k did not ex eed the a eptable 12 degrees as dened as
safety riteria.
The ontrol a tivity is evaluated under moderate turbulen e onditions ( 
= 1.5 m/s). The response is shown in Figure 26.21 for nominal onditions.
The mean (absolute) throttle rate is 23% of the maximum rate (0.36 and 1.6
degrees/s), while the mean (absolute) tailplane rate is 21% of its maximum
rate (3.2 and 15 degrees/s). The throttle a tivity ex eeds the allowed 15%.
The problem is that the gain around the set-point was in reased to obtain
less overshoot in altitude responses (Figure 26.11). However, the mean value
of tailplane rate satises the allowed 33%.

tailplane () and throttle () Ailerons() and rudder()

10 10

5 5

0 0

5 5

10 10
0 50 100 0 50 100

Figure 26.21: A tuator responses moderate turbulen e (degrees).

412
26.5.2 Analysis of lateral ontroller
The performan e and robustness of the lateral ontroller is also evaluated by
a series of time responses under the onditions of Table 26.8. The rst test
is to follow the lateral displa ement of 50m. The time responses are plotted
in Figure 26.22. Both requirements on rise time and overshoot are satised.
Noti e that the sideslip angle is redu ed to zero.

lateral deviation () heading(),sideslip()

60 6

40 4

20 2

0 0

20 2
0 20 40 60 0 20 40 60

Figure 26.22: Time responses lateral displa ement 50 m.

The heading response is tested by a heading ommand of 5 degrees. The step

responses are plotted in Figure 26.23. The rise and settling time requirements
are satised. However, the overshoot is about 20% in order to make the lateral
deviation from the ight path zero.

lateral deviation () heading(),sideslip()

10 8

6
0
4
10
2
20
0

30 2
0 20 40 60 0 20 40 60

Figure 26.23: Time responses heading 5 degrees.

In ase of a o-ordinated turn, the roll angle must not be larger than 30
degrees. In Figure 26.24, the responses for a turn of 90 degrees in 30s is shown.
The roll angles are limited to 30 degrees. The lateral displa ement is dierent
for low and high airspeed, but onverges to zero in all ases.
The responses for engine failure are plotted in Figure 26.25. Under all
onditions, the air raft is stabilized in about 40s. The requirements are fullled
ex ept for the large overshoot of roll angle when the engine is restarted. The
slightly os illatory responses belong to the air raft with large mass. Noti e the
ee tive, aggressive rudder ommands, and the zero sideslip responses.
Ride quality in ludes lateral a eleration and damping. The lateral a el-
eration initially had large peak values for normal manoeuvres. The lateral
a elerations are mainly due to sudden roll angle hanges, like in o-ordinated
turns. In order to redu e the lateral a eleration, a feedba k is added in the
inner loop ontroller from roll angle to rudder dee tion. This redu es the

413
 roll angle ()
40 lateral deviation () heading(),sideslip()
100 150

20 0 100

100 50
0
200 0

20 300 50
0 20 40 60 0 20 40 60 0 20 40 60

Figure 26.24: Time responses o-ordinated turn 3 degrees/s for 30s.

lateral deviation () heading(),sideslip()
10 4

0
2
10
0
20
2
30

40 4
0 50 100 150 200 0 50 100 150 200

roll angle () Ailerons(),rudder()

10 20

10
5
0
0
10
5
20

10 30
0 50 100 150 200 0 50 100 150 200

Figure 26.25: Time responses in ase of engine failure and restart.

lateral a elerations, but for a o-ordinated turn with heading rate of 3 de-
grees/s, the peak value is still 0.04g. During the simulated heading and lateral
manoeuvres, the a eleration is smaller than the allowed 0.02g. During the en-
gine failure, the maximum lateral a eleration is about 0.08g, whi h is smaller
than the allowed 0.2g. The se ond riterion in ludes good damping. Responses
to lateral displa ements are ni ely damped with no overshoot. Heading angle
responses have too mu h overshoot; priority is given to rise time and settling
time requirements at the low altitude.
For safety reasons, the roll angle is limited by 30 degrees. This was a hieved
by the bounds of the FLC for roll angle (see simulation of o-ordinated turn).
The se ond riterion in ludes sideslip angle. Sideslip angle is de reased to zero
at all times, even in ase of an engine failure, by using an integrator. Moreover,
the RMS of sideslip angle is 0.52 degrees in open-loop, and slightly smaller, 0.49
degrees, in losed loop for unit RMS intensity lateral Dryden gust. The RMS
values of heading angle are 0.58 and 0.29 degrees, respe tively.
The lateral ontrol a tivity is evaluated under moderate turbulen e on-
ditions (  = 1.5m/s). The responses are shown in Figure 26.21. The mean
aileron rate and the mean rudder rate are 8% (2.1 and 25 degrees/s) and 6%

414
 of their maximum values (1.5 and. 25 degrees/s), respe tively. The ontrol
a tivity riterion of 33% is met. It is noti ed that the roll angle response under
moderate turbulen e onditions is smaller than the allowed 5 degrees.

26.6 Results of the Automated Evaluation Pro-

edure
Finally, the ontroller is tested by the automated evaluation pro edure under
varying enter of gravity and time delay (four test ases, see se tion 14.3).
During the rst traje tory segment, an engine failure (a) and restart (b) takes
pla e. In Figure 26.26, the (nearly identi al) lateral displa ements are shown.
The steady-state value of lateral deviation does not onverge to zero. On the
other hand, roll angle and lateral a eleration are limited and sideslip angle
is redu ed to zero (see also Figure 26.25). The se ond segment involves a o-

First segment: top view Second segment: lateral deviations

300

100
200

50
lateral deviation [m]

100
xdeviation [m]

0 0
0 a b 1
1 c d 2

100
50

200
100

300
20 18 16 14 12 10 8 6 4 2 0 1 2 3 4 5 6 7 8
yposition (YE) [km] along track distance from point 1 [km]

Figure 26.26: Lateral responses Segment I and II.

ordinated turn of 3 degrees/s. In Figure 26.26, the lateral deviation is shown.

The roll angle and lateral deviation are within the bounds, but the peak of
lateral a eleration is slightly too high (0.04g) (see also Figure 26.24).
During the third segment, a landing approa h of -6 and -3 degrees is simu-
lated. In Figure 26.27, the longitudinal responses of the air raft are plotted; all
are within the bounds that spe ify a eptable behaviour. Verti al a eleration
is during the approa h of -6 degrees slightly too large. Finally, the ee t of a
windshear model is onsidered. In Figure 26.27, the verti al deviations from the
desired path are shown. The responses are just inside the predened bounds.
This response may be improved by in orporating forward-looking (pilot-like)
guidan e strategies in ase of windshear. For example, extra throttle setting
at the en ounter of the windshear, in order to ompensate the energy loss due
to the mi roburst and outow of the windshear. Implementing su h rules into
the FLC ould be performed in a similar way as for engine failure.
The numeri al results based on the simulations are given in Table 26.9. The
values indi ate that all spe i ations are satised (smaller than 1) ex ept the

415
 Third segment: altitude deviations Fourth segment: altitude deviations
30 30

20 20

altitude deviation [m]

3
altitude deviation [m]

10 2 f 3 10
4
g h
0 0

10 e 10

20 20

30 30
16 15 14 13 12 11 11 10 9 8 7 6 5 4 3 2 1
xposition (XE) [km] xposition (XE) [km]

Figure 26.27: Longitudinal responses Segment III and IV.

omfort riteria in ase of o-ordinated turn and glide-slope apture. These

problems were already dis ussed in detail in the previous se tions. The best
solution is by modi ations of the inner loop ontrollers, sin e the a elerations
are proportional to attitude rates.

Table 26.9: Numeri al results of the evaluation pro edure

Segm. I Segm. II Segm. III Segm. IV Total

Performan e 0.1332 0.1582 0.1593 0.5086 0.2398
Perf. Dev. 0.0319 0.1090 0.0671 0.1574 0.0914
Comfort 0.4078 1.9181 1.0779 0.5829 0.9967
Safety 0.0082 0.0205 0.0078 0.2368 0.0683
Power 0.0035 0.0069 0.0152 0.0293 0.0137

26.7 Con lusions

Be ause the ight ontrol problem is relatively well-understood in general, pilot
heuristi s are available and linearized air raft models an be derived. Therefore,
a hybrid ontroller stru ture is proposed. The inner loop ontrollers onsist of
lassi al attitude ontrollers whi h are tuned using the air raft models. For
outer loop (tra king) ontrol, longitudinal and lateral Fuzzy Logi Controllers
(FLCs) are developed using pilot heuristi s of ying an air raft (Figures 26.5
and 26.8). The on lusions of the FLC approa h are, regarding:

 The Methodology: Fuzzy logi provides a transparent interfa e between

the low-level, attitude ontrol of air raft and high-level reasoning of hu-
man pilots. Experien ed human pilots know how to handle an air raft
satisfa torily, i.e. they know how to ombine ompeting riteria like per-
forman e, robustness and passenger omfort without mathemati al for-
mulations. The pilot heuristi s are aptured through the use of ifthen

416
 rules and linguisti terms like small, big, et . for ea h variable (Chap-
ter 13). Additional rules for low airspeed and engine failure are in luded
as well, whi h shows that gain s heduling for dierent ight onditions
and ex eption handling an be readily in orporated.

 Tuning and E ien y: The translation of human heuristi s into a FLC

is a systemati pro edure [248, but ne-tuning (optimizing) a fuzzy logi
ontroller based on time responses is a matter of trial-and-error like in
lassi al ontrol. However, using the provided guidelines of se tion 26.3,
a basi understanding of the inuen e of ontroller parameters (Chap-
ter 13), and following the three-step pro edure of se tion 26.4, a satis-
fa tory ontroller an be obtained. Our experien e is that the design
involves a low number of iterations, ertainly when knowledge of lassi al
ontrol theory and ight me hani s is present.

 User-friendliness: In order to design a FLC, basi knowledge of fuzzy

logi theory and lassi al ontrol is ne essary. Our experien e is that
fuzzy logi ontrol is not very di ult to learn, with a learning urve
omparable to lassi al ontrol.

 General Results: The ontroller is evaluated with respe t to performan e,

robustness, ride quality, safety and ontrol a tivity riteria (se tions 26.5
and 26.6). The results show that a ompromise is found in whi h per-
forman e and robustness properties are good, at the ost of slightly too
high verti al and lateral a elerations.

 Controller omplexity: The obtained FLC is of low order, deterministi ,

and transparent. The ontroller stru tures are based on human ontrol
strategies whi h are very visible (re ognizable). The interpretation is
easy be ause of the rule based hara teristi s, with lear linguisti labels
like big and small a tions. Even for non- ontrol-engineers, the a tions
of the FLC are easy to understand. The nonlinear, stati fun tions of
the FLCs from error signals to ontrol ommands an be implemented by
look-up tables in order to simplify al ulation, validation and erti ation
(Chapter 13).

 Lessons learned: Not being experts in ight ontrol and aeronauti s, the
design involved some extra time to a hieve a realisti ontroller. How-
ever, our experien e shows that FLC, ontroversial but applied in many
te hni al and non-te hni al areas [136, has potential for ight ontrol
problems.

417
418
 Part III

HIRM part

419
27. The HIRM Design Challenge
Problem Des ription

Ewan Muir - 1 2

Abstra t. A ommon ontrol law design problem has been spe i-

ed based on a military ai raft onguration as part of GARTEUR
A tion Group AG-08 on Robust Flight Control. The air raft
mathemati al model des ription is a ompanied by a set of design
spe i ations for a wide ight envelope ontrol law. The design
spe i ations are mirrored by a set of evaluation riteria to fa ili-
tate omparison of dierent design methods. The design hallenge
do ument also presents parti ipants with a list of attributes, ov-
ering the design method, design pro ess and resultant ontroller,
whi h they should report on to provide a balan ed assessment of
the omplete development of their ontrol law in addition to a pre-
sentation of results.

27.1 Introdu tion

27.1.1 Ba kground
As many of the parti ipants in the GARTEUR a tion group are primarily in-
terested in military air raft, it was felt that a ontrol design task based around
a military air raft should be spe ied in addition to the ivil RCAM prob-
lem des ribed in hapter 14. The di ulty with spe ifying su h a problem is
that modern military air raft mathemati al models are lassied. The military
problem for the GARTEUR group was based on the only un lassied model of
high enough delity available at the time.
The model is based on the High In iden e Resear h Model (HIRM) s aled
up to a size representative of modern ombat air raft. HIRM was originally
developed under a Defen e Resear h Agen y programme as a vehi le for high
angle of atta k aerodynami investigations. The origin of the model explains
the un onventional onguration with both anard and tailplane plus an elon-
gated nose.

1
FDS Dept, DRA Bedford, Bedford MK41 6AE, UK
2
The following authors ontributed to the original HIRM design denition: Mr S Bennanni
(DUT), Dr R Hyde (CCL), Mr P Lambre hts (NLR), Mr D Moormann (DLR), Mr S S ala
(CIRA), Mr J S huring (NLR), Mr J Terlouw (NLR). Signi ant ontributions to the design
riteria were also made by Dr J Irving (BAe-MA) and Mrs K Sthl-Gunnarsson (SMA).

421
 The main obje tive of the study is the design of a ontrol augmentation
system for HIRM. The ight ontrol system is to give good handling quali-
ties a ross the spe ied ight envelope and also provide robustness to unmod-
elled plant dynami s, modelling un ertainties and variations in operating point
within the ight envelope. A eptable noise and disturban e reje tion must
also be demonstrated. The problem is written to provide an aerospa e ontrol
law ben hmark, whi h is both on ise and bounded, with a set of representative
design aims for both robustness and handling qualities. At the same time, it
ontains most of the elements ae ting the implementation of ontrol laws in
real systems and whi h designers need to allow for.

27.1.2 Contents of hapter

This hapter summarises the full HIRM model and problem spe i ation de-
s ribed in [177. Se tion 27.2 of this hapter details the HIRM onguration
and the model stru ture and omponents. Se tion 27.3 des ribes the ontrol law
design problem and design spe i ations. The riteria used in the evaluation
of the ontrol laws are outlined in se tion 27.4.

27.2 Des ription of the HIRM model

27.2.1 Introdu tion
The High In iden e Resear h Model (HIRM) is a mathemati al model of a
generi ghter air raft whose onguration is shown in Fig. 27.1. This mathe-
mati al model is based on aerodynami data obtained from wind tunnel tests
and ight testing of an unpowered, s aled drop model. The aerodynami s
therefore ontain degrees of non-linearity representative of modern ombat air-
raft. The wind tunnel and drop models were originally designed to investigate
ight at high angles of atta k whi h is why the dataset extends over a wide an-
gle of atta k and sideslip range ( 50 to +120, 50 ) but does not in lude
ompressibility ee ts resulting from high subsoni speeds.
The HIRM mathemati al model used for this design study was derived by
s aling up the data obtained from the drop model to reate an air raft of F-18
proportions. Engine, sensor and a tuator models have been added to reate a
representative, non-linear simulation of a twin engined, modern ghter air raft.
This gives an air raft with the hara teristi s shown in Table 27.2.1.
The air raft is basi ally stable both longitudinally and dire tionally. There are
however ombinations of angle of atta k and ontrol surfa e dee tion whi h
ause the air raft to be unstable longitudinally and/or dire tionally.

27.2.2 Blo k diagram of the system

To provide a model in a ommon format for ea h of the parti ipants in the
design hallenge, the software for the six degree of freedom HIRM model was
generated automati ally by Dymola, where ea h identiable omponent of the

422
 Figure 27.1: HIRM onguration.

model was oded in the form of equations with the onne tions between those
obje ts representing their physi al intera tion.

From the physi al des ription set up in Dymola, a onsistent symboli al

mathemati al model was built automati ally by the Dymola symboli equa-
tions handler, and from that, e ient simulation ode for dierent simulation
environments was generated.

For the Matlab/Simulink simulation environment, ode was generated in

the form of a Matlab m-le and of mex-les for Fortran or C. Fortran or C-
ode ould also be generated in the neutral DSblo k format whi h an be used
dire tly within the ANDECS simulation environment.

The Simulink blo k diagram of the six degree of freedom nonlinear High
In iden e Resear h Model in luding nonlinear a tuator and sensor models is
given in Figure 27.2.

Ea h box in this blo k diagram will be des ribed in detail. In 27.2.3,

an analyti al des ription of the air raft dynami s is given. In 27.2.4, the
a tuator model dynami s are detailed and in 27.2.5, the sensor models are
des ribed. The sti k hara teristi s assumed for the model are given in 27.2.6.
In 27.2.7, the analyti al models of wind disturban es are presented.

423
 Chara teristi Magnitude Unit
mass 15296.0 kg
Ix 24539.0 kg m2
Iy 163280.0 kg m 2
Iz 183110.0 kg m
2
Ixz -3124.0 kg m2
Ixy and I yz 0.0 kg m
2
Wing area 37.160 m2
Mean aerodynami hord 3.511 m
Wing span 11.400 m

Table 27.1: Air raft model inputs denition

Before starting a
simulation
symmetrical taileron (dtsc) load ini****.mat
0 of desired flight condition time
U0 in the workspace
differential taileron (dtdc) Clock Time
trim_inputs Double click here
0 to actuator for more information
symmetrical canard (dcsc) hirmsim
outputs for simulation
0
differential canard (dcdc) Mux +
+ hirmexa hirmexs hirmy
Mux hirmex Demux
0 measured variables
Sum_inputs1 Actuator Mux Demux Sensor
rudder (drc) model HIRM model model
(16 states)
0 wind
nose suction (suctionc) wind_input

left engine throttle (thrott1c)

right engine throttle (thrott2c) CONTROLLER

Mux1

Figure 27.2: Simulink blo k diagram of HIRM air raft system

27.2.3 Air raft dynami s model

This se tion des ribes the HIRM dynami s model orresponding to the blo k
hirmex in Figure 27.2. The dynami s obje ts are depi ted in Figure 27.3.
These obje ts are:

 body des ribes the body dierential equations of motion. (see 27.2.3)

 two transformation obje ts des ribe the oordinate transformation be-

tween the body-xed oordinates of the body obje t and the geodeti
oordinates of the gravity obje t, and between the body-xed oordi-
nates of body and the geodeti oordinates of wind, respe tively. (see
27.2.3)

 al airspeed des ribes the relationship between the inertial movement, the
wind, and the movement relative to the air.

 aerodynami s des ribes the aerodynami for es and moments. (see 27.2.3)

 engine1 and engine2 des ribe the relevant engine behaviour. (see 27.2.3)

 atmosphere des ribes the model of the atmosphere. (see 27.2.3)

424
 atmosphereConst
( as a function of height )

control
aerodynamic
inputs

calcairspeed
(calculate airspeed)
outputs
engine_1
for system
body
analysis
Equations of motion
engine_2
for control
BodyFixed BodyFixed

Transfor- mation Transfor- mation

HIRM
Geodetic Geodetic
dynamics model

gravitationConst wind
( g = 9.81 m/s 2 )
wind
inputs

Figure 27.3: Dynami s obje ts of HIRM air raft model inside blo k hirmex of
gure 2.2

 gravity des ribes the gravitational inuen e (see 27.2.3)

Model inputs, states, outputs and parameters

The following inputs are available for ontrolling HIRM:

 Symmetri al and dierential taileron dee tion

 Symmetri al and dierential anard dee tion

 rudder dee tion

 nose su tion (this was a feature in the original HIRM and was retained in
this model for ompleteness but it was not used for the purposes of this
design exer ise)

 left and right throttle settings (these are used symmetri ally)

Be ause the vehi le started as a large s ale model for wind tunnel and drop
testing, HIRM's wings were too slender to a ommodate aileron a tuators and
so dierential tailplane and anard dee tions are used for roll ontrol.
In addition to the ontrol inputs, wind omponents in the 3 body axes an
be in luded.

425
 The air raft states onsist of:

 the 3 body axis velo ities

 the 3 body axis angular rates

 the 3 air raft attitudes

 the 3 omponents of air raft position

 4 engine states (ea h engine is represented by a 2nd order model)

Control law designers an sele t from the following list of outputs as feed-
ba k signals:

 the 3 body axis angular rates

 the 3 air raft attitudes

 the 3 body axis a elerations

 air speed

 Ma h number

 altitude

 angle of atta k

 sideslip

In addition to these measured outputs, other signals are available for mon-
itoring purposes:

 ight path angle

 ground speed

 earth axis position North and East

 thrust from ea h engine

Rigid body Equations of motion

The equations of motion and transformations between the various axis systems
used in the model are standard. As they an be found in any textbook in ight
me hani s su h as [74, they are not des ribed in detail here.

426
Aerodynami s
The aerodynami for e and moment oe ients are given by the summation of
several omponents. Most omponents have the form C ( ; d)
ab . The oe ient
for for e or moment a with respe t to parameter b is determined by linearly
interpolating between the values given in a look-up table as a fun tion of the
variables and d. Variable is usually angle of atta k and variable d either
sideslip or relevant ontrol surfa e dee tion.
Be ause of a dis ontinuity in the data at 20 , it was suggested that
designers should not evaluate their ontrol laws at this angle of atta k.

Engine model
Ea h engine is modelled as shown in gure 27.4.

Throttle Thrust Sea Thrust

demand demand level at altitude
thrust
Throttle Rate 1
non-linearity + 1+0.07958s limiter s 0
-

Figure 27.4: Engine model

A throttle demand of 0 sele ts idle whi h is 10 kN of thrust. A throttle

demand of 1 orresponds to a maximum dry thrust of 47 kN. Full reheat is
sele ted when the throttle demand equals 2 at whi h time the thrust is 72 kN.
The rate at whi h the thrust hanges is dependent on whether the engine is in
dry thrust or reheat. For dry thrust, the maximum rate of hange of thrust is
12 kN/s whereas in reheat it is 25 kN/s.
If the rate limit is ignored, the engine dynami s an be represented by a
se ond order transfer fun tion with a damping of unity and a natural frequen y
of 2 rad/s.
The sea level engine thrust FE0 is s aled with density as follows:


FE = FE0 :
0
The engine setting angles are zero and so the thrust a ts parallel to the
air raft x-body axis. Therefore:

FEx = FE ; FEy = 0; FEz = 0

Atmosphere
The atmospheri model is that of a standard atmosphere with the following
pressure and temperature variations.

427
 T = T0 Tgradh
8   R Tg
<
p0 TT0 grad
h < 11000 m
p =
h  11000 m
: g (h 11000)
p0 exp R T0

0:0065 K=m h < 11000 m
Tgrad =
0:0 K=m h  11000 m

288:15 K h < 11000 m
T0 =
216:65 K h  11000 m

101325:0 Pa h < 11000 m
p0 =
22632:0 Pa h  11000 m :
In the equations above, p is the stati pressure, T the absolute temperature
and Tgrad its rate of hange w.r.t. height. The subs ript 0 for temperature
and pressure represents the starting point for the interpolation, for heights of
either h = 0 m or h = 11000 m.

Gravity
The gravitational onstant is assumed to be invariant with altitude and has a
value of g = 9.80665 m/s .
2

27.2.4 A tuator models

Available motivators
The motivators available for use in ontrolling the air raft are listed in se tion
27.2.3. The a tuator representations for ea h aerodynami ontrol surfa e are
des ribed in se tions 27.2.4 to 27.2.4 below.

Taileron a tuator
The taileron onsists out of a starboard taileron and a port taileron, whi h an
be ontrolled independently.
The taileron a tuator is modelled by the third order transfer fun tion

1
(1 + 0:026s)(1 + 0:007692s + 0:00005917s2)
with a 80 deg/s rate limit.
Dee tion limits for starboard and port taileron are +10 deg to 40 deg.
A positive dee tion is dened as trailing edge down.
The inputs of the taileron a tuator model are dierential and symmetri al
tailplane demand.

428
Rudder a tuator
The rudder a tuator is represented by a se ond order transfer fun tion with a
80 deg /s rate limit:
1
(1 + 0:0191401s + 0:000192367s2)
Dee tion limits are 30 deg. A positive dee tion is dened as trailing
edge to port.

Canard a tuator
The anard onsists out of a starboard anard and a port anard, whi h an
be ontrolled independently. The following transfer fun tion is valid for both
parts of the anard.
The anard a tuator is represented by a se ond order transfer fun tion with
a 80=s rate limit:
1
(1 + 0:0157333s + 0:00017778s2)
Dee tion limits for starboard and port anard are +10 deg to 20 deg. A
positive dee tion is dened as trailing edge down.
The inputs of the anard a tuator model are dierential and symmetri al
anard demand.

27.2.5 Sensor models

The list of sensor signals available for ontrol feedba k purposes is given in
se tion 27.2.3. Two dierent types of sensor dynami s models are used for the
air data and attitudes, and the rates and a elerations, as shown in Fig. 27.5

Sensor dynami s and signal onditioning

The transfer fun tions for the individual omponents are as follows (assuming
a Flight Control Computer (FCC) operating at 80Hz):
 air data sensor dynami s:

1=(1 + 0:02s)

 attitude sensor dynami s:

1=(1 + 0:0323s + 0:00104s2)

 anti-aliasing lter:

1=(1 + 0:00398s + 0:0000158s2)

429
 Air data and attitudes

Sensors FCC Actuators

Sensor Anti-aliasing D/A
compute
dynamics filters conversion
delays

Rates and accelerations

Sensors Sensor FCC Actuators

Anti-aliasing compute D/A
dynamics &
filters delays conversion
notch filters

Figure 27.5: Sensor hardware assumptions

 rate and a eleration sensors ( ombined sensor and not h lter):

(1 0:005346s + 0:0001903s2)=(1 + 0:03082s + 0:0004942s2)

 averaging on rate and a eleration data:

(1 0:00208s)=(1 + 0:00417s)
 ompute delay:
(1 0:0062s)=(1 + 0:0062s)
 D/A onverter:
(1 0:00208s)=(1 + 0:00417s)

Simplied sensor dynami s

The sensor dynami s from 27.2.5 an be approximated by the following trans-
fer fun tions:

 Rates and a elerations:

(1 0:0173s + 0:00019s2)=(1 + 0:0401s + 0:000704s2)

 Air data:

(905:92 14:437s + 0:116s2)=(908:77 + 29:573s + s2 )

 Attitudes:

(7161:8 82:317s + 0:3417s2)=(7162:3 + 190:85s + s2 )

Designers were instru ted that the approximations above ould be used
during the design phase, but the full sensor dynami s des riptions from 27.2.5
were to be used for ontrol law analysis.

430
Sensor noise
The measurement noise is generated by passing pink noise of unit rms power
through the following lters and adding this to the feedba k signal.

 Noise hara teristi s for angular rates p, q, r, air speed, angle of atta k
and sideslip:

  
0:05 0:053s
1 + 0:0089s + 0:000041s2 1 + 0:053s

 Noise hara teristi s for linear a elerations ax , ay and az :

1 + 0:0039s + 0:000078s2
   
0:05 0:053s
1 + 0:011s + 0:000063s2 1 + 0:053s 1 + 0:0018s + 0:000078s2

 It is assumed that the altitude, Ma h number and attitudes ,  and

signals are relatively noise free. However these signals are not be dier-
entiated around the ross-over frequen y, as it is assumed that su h a
dierentiated signal is more noisy than the orresponding rate measure-
ment.

Measurement errors
The measurements listed in se tion 27.2.3 above an be assumed to be a urate
ex ept for the following:

 and  : to within 2 deg.

These errors are assumed onstant during the period of a simulation.

27.2.6 Sti k hara teristi s

The following sti k hara teristi s are to be used during the ontrol law evalu-
ation against the performan e riteria of 27.3.5.

In eptor Dee tion amplitude Sti k for es

Longitudinal sti k -72 to +120 mm 1.2 N/mm
Lateral sti k -80 to +80 mm 0.5 N/mm

Table 27.2: Sti k hara teristi s.

27.2.7 Atmospheri turbulen e models

The turbulen e model used is the Dryden spe tral model des ribed in [2.

431
27.3 Control problem denition
27.3.1 Introdu tion
Modern ghter air raft are designed with either unstable or only marginally
stable ongurations whi h ne essitate ontrol augmentation systems. The
obje tive of the HIRM design hallenge is to design a ontrol augmentation
system whi h will tra k the demands listed in se tion 27.3.2 with a response
whi h is in keeping with the handling qualities listed in se tion 27.3.5 a ross the
ight envelope dened in se tion 27.3.3. They should also demonstrate good
disturban e reje tion apabilities and insensitivity to sensor noise.

27.3.2 Control strategy

Pilot ommands
The pilot ommands should ontrol the following responses.

 lateral sti k dee tion should demand velo ity ve tor roll rate. The ve-
lo ity ve tor roll is a roll performed at onstant angle of atta k and zero
sideslip. The velo ity ve tor roll will therefore vary from a pure body-axis
roll rate at 0 deg angle of atta k to pure body-axis yaw rate at 90.
 longitudinal sti k dee tion should demand pit h rate.

 rudder pedal dee tion should demand sideslip.

 the throttle lever should ontrol velo ity ve tor air speed.

Requirements of the ontrol system

The automati ontrol system should satisfy the following requirements

 The following ross- ouplings should be minimised. Changes in air speed

should not ause pit h transients. Pit h transients due to rolling ma-
noeuvres should be minimised. Velo ity ve tor rolls should ause minimal
sideslip ex ursions.

 The pit h rate demand system should limit at 10 deg and +30 deg angle
of atta k and 3g and +7g normal a eleration. Overshoots of 5 deg and
0:5g are allowable on angle of atta k and normal a eleration respe tively.
The overshoots must be washed out and the air raft should return to the
limiting values within 2 se onds.

 The ontrol system must make use of the motivators in an e ient man-
ner. For example, the anards and tailerons must trim the air raft in a
manner whi h minimises drag.

432
27.3.3 Robustness onsiderations
The ontrol system should maintain its good ying qualities and robustness
a ross the ight envelope dened in se tion 27.3.3. In addition, it should
demonstrate a toleran e to the model un ertainties listed in se tion 27.3.3, the
measurement errors given in se tion 27.3.3 and omplian e with the hardware
implementation issues identied in se tion 27.3.3.

Design envelope
The design envelope for the HIRM ontrol law is

 Ma h 0:15 to 0:5,
 Angle of atta k 10 deg to 30 deg,
 Sideslip 10 deg,
 Altitude 100 to 20000 ft.

Modelling errors
For linear assessments, the ontrol laws need to be robust to the following errors
in the aerodynami moment derivatives

Cmw 0:001
Clv 0:01
Cnv 0:002
Cmq ; Clp ; Clr ; Cnp ; Cnr 10%
CmT S ; CmCS ; ClT D ; ClCD ; ClRUDDER ; CnT D ; CnCD ; CnRUDDER 10%
For nonlinear assessments, the ontrol laws need to be robust to the fol-
lowing errors in the total moment oe ients

Cm 0:03
Cl and Cn 0:008
The engine, a tuator and sensor models an be assumed to be a urate and
have zero toleran es.

Measurement errors
Sensed values of angle of atta k and sideslip may not be a urate. The ontrol
system must be robust to the following measurement errors:

 and  2 deg

433
Hardware implementation onsiderations
The ontrol laws must be designed to operate at 80Hz, the iteration rate of the
ight ontrol omputer (FCC). They an be designed by negle ting hardware
implementation issues, but must be robust to the dynami s of stru tural lters,
D/A onverters and omputational delay, whi h are given in se tion 27.2.5
above.

27.3.4 Robustness requirements

In the multi-loop ase, the losed-loop system should be able to withstand the
appli ation of simultaneous and independent gain and phase osets at the input
of ea h one of the a tuators as shown in Fig. 27.6 without be oming unstable
(note that the losed-loop system shown in Fig. 27.6 is only an example and
does not represent a mandated ontrol system stru ture). The orresponding
perturbation matrix P will be of the form

P = diag(K1 e |1 ; : : : ; K6 e |6 )

with Ki and i taking values in the regions shown in Fig. 27.7 (no toleran es)
and Fig. 27.8 (toleran es applied).

Add additional gain and phase or

break for open loop analysis here

Pilot
demands e u Outputs
Command
Actuators Aircraft
path filtering +
-

u = actuator demands
e = error signals Controller Sensors

Figure 27.6: Closed-loop system showing point for analysis

27.3.5 Performan e requirements

1. For single loop analysis, the open-loop Ni hols plot of the frequen y re-
sponse between ea h a tuator demand u and the orresponding error
signal e, obtained by breaking the loop at the point shown in Fig. 27.6
while leaving the other loops losed, should avoid the regions shown in
Fig. 27.9 (no toleran es applied) and Fig. 27.10 (toleran es applied). The
toleran es are listed in se tion 27.3.3 and should be applied to the lin-
earised air raft models. Note that when performing the frequen y re-
sponse, a gain of -1 needs to be in luded on the input or output to obtain

434
 the orre t phase response. These hara teristi s should be valid for ea h
loop.

2. For tests 2 and 3, the results from non-linear frequen y responses should
be used. These responses are obtained by supplying a sinusoidal input
of in reasing frequen y to the system. The frequen y should in rease
logarithmi ally from 0.5 to 20 rad/s over a 20 s period. Sti k for es of
3.24 and 9.71 pounds (14.4 N, 12 mm and 43.2 N, 36 mm) should be used
in pit h and 0.9 and 2.7 pounds (4 N, 8mm and 12 N, 24 mm) in roll.
The time response of the pit h or roll attitude should be analysed using a
Fast Fourier Transform to obtain the gain and phase hara teristi s of the
response. The gain and phase an then be plotted in gures 27.11, 27.13
and 27.14. To ensure PIO resistan e, the hara teristi s of the frequen y
response between sti k for e in pounds ( 1 pound for e = 4:448N) and
pit h or bank attitude in degrees should meet the following requirements.

 The pit h and bank attitude absolute amplitude gains at 180 deg
phase should be < 16dB.
 The magnitude of the average phase rate, _ average , dened as

j_ averagej = 2f f f

;

should lie either within the level 1 or level 1 boundaries shown in
Fig. 27.12. It should be noted that the plot in Fig. 27.11 is only an
example.

3. This riterion omplements the handling qualities metri s illustrated in

gures 27.11 and 27.12 [3. The non-linear frequen y response from longi-
tudinal sti k for e to pit h attitude should lie within the level 1 boundary
of Fig. 27.13. The non-linear frequen y response from lateral sti k for e
to roll angle should lie within the region labelled as providing a good
response in Fig. 27.14. The frequen y responses are non-linear and are
between the pilot demand and the air raft pit h attitude or roll angle (i.e.
the integral of the velo ity ve tor roll rate). For these riteria, the term
open-loop means that there is no pilot losing the loop between air raft
response and pilot demand. The responses should be obtained for dier-
ing sti k for es: pit h sti k for es of 3:24 and 9:71 pounds (14.4 N, 12
mm and 43.2 N, 36 mm) and roll sti k for es of 0:9 and 2:7 pounds
(4 N, 8 mm and 12 N, 24 mm) should be used.

4. The ratio of dropba k db to steady state pit h rate qss should be

0 < db =qss < 0:25s

for pre ision tra king. Dropba k is dened in Fig. 27.15 [3.

435
 5. The peak pit h a eleration in response to a step input should be a hieved
in < 0:15 s.
6. The roll mode time onstant should be  0:4s.
7. The maximum roll a eleration in response to a step input should be
< 600 deg/s2.
8. The maximum roll rate should be approximately 70 deg/s.
9. Sideslip response requirements. The oupling in sideslip due to roll should
be minimised and not ex eed 0:5 deg for  < 15 deg and 2 deg for
 > 15 deg. The step response to sideslip demand should lie within the
boundaries shown in Fig. 27.16. The sideslip response should also have
an a eptable level of damping ( > 0:5).
10. Speed ontrol requirements. The speed response should have minimal
overshoot ( < 3%). The time onstant for small amplitude speed demands
< 1%
( throttle travel) should be in the range 0:75 to 1:5 se onds. For
large amplitude speed demands, maximum use should be made of the
engine performan e.

11. Avoidan e of stru tural oupling. To avoid airframe/FCS stru tural ou-
pling the following limitations should be observed.

 The maximum high frequen y (frequen ies above 4 Hz) gain from
pit h rate (rad/s) to ontrol surfa e dee tion (rad) should be < 3:0.
 The maximum high frequen y (frequen ies above 4 Hz) gain from
normal a eleration (g) to ontrol surfa e dee tion (rad) should
be < 0:09 rad/g.
12. Disturban e reje tion. The ontrol system should minimise the ee t
of atmospheri disturban es on the air raft's ight path. The ee t of
turbulen e on the air raft should be assessed during straight and level
ight.

27.3.6 S heduling onsiderations

1. The amount of s heduling should be minimised.

2. Any s heduling must be against measurable air raft states as dened in

se tion 27.2.3 above.

3. If angle of atta k is used for s heduling, any linear analysis must take
into a ount the impli it feedba k whi h this generates.

436
27.4 Evaluation riteria
The nal ontrol law design should meet the requirements spe ied in se -
tion 27.3 a ross the ight envelope. For veri ation purposes, a set of spe i
evaluation riteria are set out below, against whi h all entries to the design
hallenge should be measured.
The riteria have been divided into four sub- lasses whi h are: robustness,
performan e, physi al onsiderations and ontrol a tivity.
Most of the evaluation riteria are based on two parti ular ight onditions
whi h are:

 Flight ondition 1: Ma h 0:4, 10000 feet altitude, 8:67 degrees in iden e

and zero sideslip.

 Flight ondition 2: Ma h 0:24, 20000 feet altitude, 28:9 degrees in iden e

and zero sideslip.

The rst ight ondition an be thought of as a nominal ight ondition,

and the se ond as an edge of the envelope ondition, likely to ause greater
stability and a tuator limiting problems. The robustness assessments in se tion
27.4.1 are based on linear analysis. The remaining evaluations in se tions
27.4.2, 27.4.3 and 27.4.4 are based on nonlinear responses and analysis.

27.4.1 Robustness riteria

Nominal ight Ni hols plot
For ight onditions 1 and 2, Ni hols plots are obtained by breaking the losed-
loop at the 6 plant a tuator demands (symmetri and dierential taileron,
symmetri and dierential anard, rudder and symmetri engine), linearising
the model and performing a frequen y response. These orrespond to the
robustness test des ribed in Figure 27.9. The Ni hols plots are he ked to
ensure that they do not infringe the ex lusion zone shown in Figure 27.9.

Robustness to multivariable un ertainty

For ight ondition 1, the losed-loop system stability is he ked when the si-
multaneous gain and phase osets des ribed in 27.3.4 and shown in gure 27.7
are added. Identi al gain and phase osets are added to ea h of the loops. As a
minimum, the gain and phase osets at the 4 orners of the plot on gure 27.7
are tested.

Robustness to parametri un ertainty

The test in se tion 27.4.1 is repeated for ight onditions 1 and 2 with the
following perturbations added to the linear model aerodynami oe ients as
des ribed in se tion 27.3.3

437
 Cmw 0:001
Clv 0:01
Cnv 0:002
Cmq ; Clp ; Cnr 10%
Clr ; Cnp +10%
CmT S ; CmCS ; ClT D ; ClCD ; ClRUDDER ; CnT D ; CnCD ; CnRUDDER 10%

27.4.2 Performan e riteria

Disturban e reje tion
Using the Dryden turbulen e model with moderate turbulen e levels, the RMS
variations of pit h rate, roll rate, yaw rate and normal a eleration are to be
re orded for ight ondition 1.

Gibson riteria
For ight onditions 1 and 2, the plots orresponding to Figures 27.13 and
27.14 are onstru ted. Due to rate limiting and other non-linear ee ts, the
frequen y responses are, in pra ti e, a fun tion of sti k for e. Hen e the plots
are onstru ted using sinusoids of 3:24 and 9:71 pounds for e (14.4 N, 12
mm and 43.2 N, 36 mm) in pit h and 0:9 and 2:7 pounds for e (4 N, 8 mm
and 12 N, 24 mm) in roll.

Assessment manoeuvres
The nonlinear response of the air raft should be tested by applying the step
inputs given in Table 27.3.
Responses showing the pit h rate, velo ity ve tor roll rate, sideslip angle,
angle of atta k and air speed should be plotted for all of the above responses.
Normal a eleration should be plotted if this rea hes its positive or negative
limiting value as dened in se tion 27.3.2.
The above series of manoeuvres are repeated with the following un ertain-
ties and measurement errors:

Cm 0:030
Cl 0:008
Cn 0:008
in ombination with a measurement error of 2 deg on in iden e and +2 deg
on sideslip.

27.4.3 Physi al onsiderations

Airframe loading and departure prevention
It should be re orded whether the following riteria (see Se tion 27.3.2) are
satised:

438
 Manoeuvre Flight ondition
1 5/s pit h rate demands M = 0:2, h = 1000 ft
M = 0:3, h = 5000 ft
M = 0:5, h = 15000 ft
2 70/s roll rate demands M = 0:3, h = 5000 ft
M = 0:5, h = 15000 ft
3 10 sideslip demands M = 0:2, h = 1000 ft
M = 0:3, h = 5000 ft
M = 0:5, h = 15000 ft
4 +100 kn (w.r.t. M = 0:3) step M = 0:3, h = 5000 ft
in air speed demand
5a +10/s pit h rate demand fol- M = 0:5, h = 15000 ft
lowed by
5b 360 velo ity ve tor roll at on 30 angle of atta k limit
+70/s when air raft on 30
limit. Followed by
5 10/s pit h rate demand. Fol- when 360 velo ity ve tor roll
lowed by omplete
5d 0/s pit h rate demand when air- when air raft has unloaded to re-
raft has unloaded. sume straight and level ight ap-
proximately.

Table 27.3: Step inputs

439
  Max normal a eleration < 7g (with +0:5g maximum overshoot)
 Min normal a eleration > 3g (with 0:5g maximum overshoot)
 Max in iden e < 30 deg (with +5 deg maximum overshoot)
 Min in iden e > 10 deg (with 5 deg maximum overshoot)
 Settling time < 2s following the manoeuvre

Stru tural oupling

To avoid airframe/FCS stru tural oupling, the ontrol system should be anal-
ysed to ensure that the limitations stated in se tion 27.3.5 are observed.

27.4.4 Control a tivity riteria

Max roll rate and pit h rate
For ight ondition 2, the maximum roll and positive pit h rates obtainable
should be re orded ( orresponding to full lateral and full longitudinal sti k).
Time histories of the response and the ontrol surfa e dee tions should be
given.

Control a tivity due to noise and turbulen e

For ight ondition 1, the RMS a tuator demand a tivity, relative to the steady-
state values of the a tuators when no turbulen e is present, due to the Dryden
turbulen e model should be re orded for ea h of the six a tuator demands.
The above should be repeated but in response to the sensor noise hara -
teristi s dened in se tion 27.2.5.

440
 -4.5 6 Gain oset (dB)

-2.5

Phase

0 30 -
oset ( )

+2.5

+4.5

Figure 27.7: Gain and phase osets ex luding toleran es

6 Gain oset (dB)

-3.0

-1.0
Phase

0 30 -
oset ( )

+1.0

+3.0

Figure 27.8: Gain and phase osets in luding toleran es

441
 6 6 Gain (dB)

180 145 -
Phase ( )

-3

-6

Figure 27.9: Gain and phase ex lusion zones on Ni hols plot for single loop
analysis frequen y response ex luding toleran es.

6 Gain (dB)

4.5

1.5

180 145 -
Phase ( )

-1.5

4.5

Figure 27.10: Gain and phase ex lusion zones on Ni hols plot for single loop
analysis frequen y response in luding toleran es.

442
 10

5
Gain: degrees attitude per pound stick force (dB)

f
-5 bw

-10

-15

-20
fc

-25
2f c

-30
-220 -200 -180 -160 -140 -120 -100
Phase (degrees)

Figure 27.11: Pit h/roll sti k for e to pit h attitude/roll angle frequen y re-
sponse riterion.

150

Level 2
Average phase rate (Degs/Hz)

100

Level 1

50

Level 1*

0
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
Frequency, fc, at -180 degrees phase (Hz)

Figure 27.12: Phase rate riterion.

443
 20
(100,18)

Relative open loop amplitude (dB) 15 (80,15)

10 (75,10)

(100,6)
5
L1 (75,4)

(85,2)
0

(150,3) (110,0)

5
L1

10
(140,12)
180 160 140 120 100 80 60 40
Open loop phase (degrees)

Figure 27.13: Pit h sti k for e to pit h attitude frequen y response riterion.

0.25 db

25 0.5 db

20 1 db
1 db
15

10 3 db
3 db
6 db
Magnitude (dB)

5
Sluggish response
0 PIO activity 6 db

10 12 db

Oscillation ratcheting
15
Good response Quick jerkey response
20 db
20

25
350 300 250 200 150 100 50 0
Phase (degrees)

Figure 27.14: Roll sti k for e to roll angle frequen y response riterion.

444
 q (rad/s)

pitch attitude

qstat

qmax

dropback = db

db/qstat time (s)

time (s)

Figure 27.15: Denition of dropba k.

Sideslip command requirement

2

1.8

1.6

1.4
Normalised sideslip angle

Upper boundary
1.2

0.8

Lower boundary
0.6

0.4

0.2

0
0 1 2 3 4 5 6 7 8 9 10
Time (seconds)

Figure 27.16: Sideslip requirements.

445
28. Design via LQ Methods

Fran es o Amato 1, Massimiliano Mattei 1

2
Stefano S ala and Leopoldo Verde 2

Abstra t. In this hapter the appli ation of the LQ based te h-

niques to the HIRM design problem is onsidered. The stru ture
of the proposed ontroller is basi ally a PI a tion designed with
optimality riteria in order to a hieve the performan e and robust-
ness requirements. Moreover, the ontrol s heme is ompleted by
a nonlinear ompensation of the dynami pressure ee t, a feed-
forward ontrol a tion, some demand shaping lters and swit hing
logi based on angle of atta k and Ma h number, as well as on the
pilot demand. Most of the design requirements are easily satised;
some others need a greater eort for tuning of the ontroller design
parameters. A pro edure to tune su h parameters, formulated in
terms of an optimization problem, is proposed. Simulation results
from the Automated Evaluation Pro edure are also provided.

28.1 Introdu tion

The following ontribution to the FM(AG08) Garteur Group is the result of
a ooperation between the Flight Me hani s and Control Group of C.I.R.A.
and the System and Control Group of the Dipartimento di Informati a e Sis-
temisti a of the University of Naples Federi o II. It onsists of the appli ation
of the LQ based te hniques to the HIRM design problem. These te hniques
(see also Chapter 4), whi h are basi ally linear plant oriented, annot be ap-
plied in a straightforward way to take into a ount all the requirements given
in Chapter 27. We propose a design y le based on the introdu tion of some
auxiliary fun tions to a hieve an optimal hoi e of the design parameters of the
ontroller, namely the LQ matrix weightings.
The main hara teristi s of the proposed ontrol s heme an be summarised
as follows:

1. the nature of the ontroller is basi ally a multivariable Proportional plus

Integral (PI) ontrol law with matrix gains omputed by means of the
standard LQ te hnique;

2. the ontroller is adaptive on the basis of the pilot demand (velo ity ve tor
roll demand requires a s heduling of the PI gains);
1
Dipartimento di Informati a e Sistemisti a, Universit degli Studi di Napoli Federi o II
via Claudio 21, 80125 Napoli, Italy, Tel.+39(81)7683172, Fax+39(81)7683686
2
Centro Italiano Ri er he Aerospaziali Via Maiorise, 81043 Capua (CE), Italy
Tel.+39(823)623949, Fax+39(823)623335

446
 3. the ontroller gains are s heduled, on the basis of the Ma h number and
of the angle of atta k, so as to over the whole operating envelope of the
air raft;

4. the robustness and performan e issues are addressed by utilizing a design

y le whi h optimizes a suitable ost fun tion, over the possible weighting
matri es of the LQ (see Chapter 4).

The hapter is organized as follows. The ontroller ar hite ture hosen and
the translation of the HIRM design riteria into method dependent obje tives
are dis ussed in Se tions 28.2 and 28.3 respe tively. Se tion 28.4 is devoted
to the des ription of the design y le,swit hing while in Se tion 28.5 some
numeri al results from the automated evaluation pro edure are presented. Fi-
nally Se tion 28.6 deals with some on lusions about the overall design a tivity
arried out by the CIRA+UNAP team. A more detailed des ription of su h
a tivities an be found in the Garteur Report TP-088-26 [9, whi h is the main
referen e of the hapter.

28.2 The Sele tion of the Controller Ar hite -

ture for the HIRM Problem

p w demand
Switching Mach,
Commands
Feedforward
w Generator
Pilot + +
demand Demand Ref 1 + +
PI gains HIRM HIRM outputs
shaping + s Nonlinear
filters - with AW Compensation
+
Reference
xm Signal -
2 Generator
ym Vm

Velocity vector +
roll rate LQO Error
computation -
Detector

Figure 28.1: HIRM Control S heme

The ontrol system ar hite ture proposed for the HIRM design problem
is s hemati ally shown in Figure 28.1. From a fun tional point of view, the
ontroller an be divided into the following sub omponents:

i) a feedforward ontrol generator;

447
 ii) a referen e signal generator;

iii) a state feedba k PI a tion designed with the LQ te hnique;

iv) swit hing logi based on the pilot demand;

v) a nonlinear ompensation of the ee t due to the variation of dynami

pressure;

vi) Anti-windup a tion due to the presen e of multiple integrators;

vii) a nonlinear transformation to ompute the velo ity ve tor roll rate from
the measured variables;

viii) a set of demand shaping lters;

ix) swit hing logi based on the  and Ma h variables;

x) a Luenberger observer to dete t and eliminate the onstant measurement
errors on the  and  angles.

28.2.1 The feedforward ontrol signal and the referen e

signal generator
The feedforward ontrol signal generator, trained o-line with the pro ess, pro-
vides the ontrol ommands to keep the air raft in straight and level ight at
ea h point of the ight envelope.
Let us onsider the nonlinear equations of the air raft motion in absen e of
un ertainties and disturban es

x_ = f (x; u) (28.1)

where

x = (V   p q r   Z engF engF 1 )T (28.2a)

u = ( T S T D CS CD R T H )T (28.2b)

are the state and the input of the air raft respe tively. To obtain the straight
and level ight ommands orresponding to a desired true airspeed Vd and to
a nominal altitude hd , sin e an analyti al approa h is not possible, we have to
numeri ally solve the following optimization problem
s X
min
z
qi2 fi2 (~x(zopt ); u~(zopt )) ; (28.3)
opt
i=1;::;7
where fi is the i-th omponent of the ve tor fun tion f , qi are weighting ele-
ments, zopt = (  engF uT )T is the optimization variable, and
x~(zopt ) = ( Vd zopt (1) 0 0 0 0 0 zopt(1) 0 hd zopt (2) 0 )T
u~(zopt ) = ( zopt (3) zopt (4) zopt (5) zopt(6) zopt(7) zopt (8) )T :
The redundan y of the ontrol surfa es (presen e of anards and tailerons),
in balan ing the moments a ting on the air raft, makes the solution of the

448
problem (28.3) indeterminate. We an use this degree of freedom to sear h,
among the possible solutions, for the one whi h minimizes the drag a ting on
the airplane, so as to satisfy the requirement given in Se tion 27.3.2. This
ondition an be translated into a modied optimization problem
s X
min
z
qi2 fi2 (~x(zopt ); u~(zopt )) + qe2 engF2 ; (28.5)
opt
i=1;::;7
qe being a further weighting element.
Hen e the solution of problem (28.5) provides the ontrol ommands, as
well as the angle of atta k value, to a hieve a straight and level ight ondition
at a minimum drag.
If we perform the omputation of su h ommands for a representative num-
ber of points in the ight envelope, we an store these values in the Flight
Control Computer (FCC) and use them as a feedforward ontrol law.
Furthermore, sin e the synthesis of the feedba k ontroller is performed
using the linearized models of the air raft in the neighbourhood of the states
and inputs resulting from the above des ribed pro edure, the linearity of the
PI ontrol a tion alls for the variable y ytr , y being the a tual value of the
output variables and ytr being the value of the output variables orresponding
to the straight and level ight. The task of the referen e signal generator is
that of providing the value of ytr asso iated with the straight and level ight
ondition of the air raft at the beginning of ea h manoeuvre.

28.2.2 The PI feedba k ontrol a tion

The ore of the proposed ontrol s heme is a multivariable PI feedba k a tion.
We used the LQ ontrol te hnique to ompute the PI matrix gains following
the guidelines drawn in [10 (see also Chapter 4).
Let us onsider the linearized model of the HIRM air raft in the form

x_ = Ax + Bu u (28.6a)

y = Cx (28.6b)

orresponding to a ertain ight ondition and assume the omplete measur-

ability of the state variables (28.2a). To satisfy the requirements given in
Se tion 27.3.2 we have to synthesize a ontroller whi h allows for the regula-
tion of four output variables: the velo ity ve tor air speed, the pit h rate, the
velo ity ve tor roll rate and the sideslip angle. We introdu e also  among the
variables to be regulated for reasons to be laried in Se tion 28.2.3; therefore
y = ( V q pw   )T = Cx.
Let us now make referen e to the losed-loop s heme shown in Figure 28.1,
where we onsider the linear plant to be the HIRM linearized model. The
state-spa e realization of the integrator is

x_ i = e (28.7a)

yi = xi ; (28.7b)

449
where e=r y is the tra king error and r is the referen e signal. We have
the following losed-loop system state-spa e equation

   
x^_ = A + BCu Kp Bu Ki x^ + 0 r
0 I (28.8)

 
where x^ = xx . Equation (28.8) an be rewritten in a ompa t form as
i
 
x^_ = A^ + B^ K^ x^ + B^2 r (28.9)

where      
A^ = A 0 ; B^ = Bu and B^2 = 0I ;
C 0 0
^ = ( Kp Ki ) is a
are the state-spa e matri es of the augmented system and K
state feedba k gain a ting on su h a system.
Obviously, the design of a state feedba k on the augmented system ^ B^ )
(A;
allows to ompute, via an appropriate partitioning of the matrix K^ , the pro-
portional and integral gain matri es for the original model (A; B; C ).
As always it happens when adopting LQ based te hniques, the key point for
the design is the hoi e of the weighting matri es for the quadrati ost fun tion
related to the auxiliary system ^ B^ ), say Q^ and R^ . Indeed, our main obje tive
(A;
is to keep as low as possible the tra king error e, while maintaining the ontrol
variables within the pres ribed ranges. This means that in the quadrati ost
fun tion, dened on system ^ B^ ), the last ve states, whi h are the states of
(A;
the integrators, should be emphasized by in reasing the relative Q^ terms (the
 is an ex eption, sin e in general it does not need to
integrator of the error on
R^ matrix, a good trade-o between
be regulated). To establish the hoi e of the
performan e and ontrol a tivity must be found; in any ase we assume Q ^ and
R^ to be diagonal matri es.
The PI ontroller stru ture proposed has three appre iable properties:

i) the low order of the ompensator (only ve linear states oming from the
integrators);

ii) the simpli ity in the omputation of the gain matri es (we use the solution
of a standard LQ problem);

iii) some intrinsi robustness properties guaranteed by the LQ ontrol (see

Chapter 4).

However, there are several problems in using this te hnique whi h for ed us
to add other omponents to the ontrol s heme:

i) it is not possible to take into a ount the nonlinear nature of the plant;

ii) it is supposed to have the omplete a essibility of the state;

iii) there is no way to dire tly take into a ount all the robustness and per-
forman e spe i ations as required by the HIRM design problem;

450
 iv) due to the presen e of multiple integrators there is a potential for winding-
up, espe ially sin e some of the ontrolled variables are related through
kinemati relationships.

To over ome problem i) we introdu ed two a tions: a nonlinear s aling

of the ontrol ommands (see Se tion 28.2.5) and a swit hing logi between
dierent sets of matrix gains, s heduled with respe t to Ma h number, angle
of atta k and pilot demand.
Con erning point ii), we have omplete a essibility of the six DOF air raft
model states; the problem is the presen e of the additional states of sensors and
a tuators. By means of an extensive ampaign of simulations on the omplete
linearized model of the air raft we ould verify that the dynami s of sensors and
a tuators do not ae t the performan e of our ontroller. Hen e we propose to
negle t them in the synthesis of the ontroller and to re-introdu e them when
verifying the performan e of the losed-loop system.
A possible way to over ome problem iii) will be dis ussed in the next hap-
ter. Nevertheless, if the designer has a deep knowledge of the plant, it is not
di ult to nd the proper weighting matri es to a hieve the performan e re-
quirements. Robustness requirements, in the form given in Chapter 27, an
only be he ked a posteriori if applying the LQ te hniques in a lassi al way.
Finally, the winding-up potential due to the presen e of the multivariable
integrators has been eliminated by a te hnique based on resetting integrators.
This te hnique, dis ussed in the TP-088-26 Garteur report [9, prevents from
wind-up phenomena, at the ost of a small performan e de rease, in the pres-
en e of saturated ommands.

28.2.3 Swit hing logi on the pilot demand

The velo ity ve tor roll manoeuvre requested in Chapter 27 opens a problem in
the ontrol law design: there are oni ting a tions to be performed depending
on the pilot demand. Indeed a speed variation whi h brings the air raft from
a straight and level ight ondition to another one annot be performed at
a onstant angle of atta k; onversely, when demanding velo ity ve tor roll
rate, the angle of atta k has to be onstant as spe ied in Se tion 27.3.2. In
order to satisfy these oni ting requirements we propose a swit hing ontroller
depending on the pilot demand: we introdu e the denomination Manoeuvre
Oriented Swit hing Logi (MOSL) for the logi whi h governs the s heduling.
If the pilot demands a velo ity ve tor roll rate greater than 10 deg=se the
gains of the ontroller swit h to a new set. The new gains have to take into
a ount the regulation of , whi h means that, in hoosing the matrix weight
Q^ , the term relative to the  error integrator has to be in reased. Note that,
for demands less than 10 deg=se , it is not ne essary to swit h the ontroller
sin e there are only small os illations indu ed on .
Assume that we have found two optimal ontrollers: one for the velo ity
ve tor roll manoeuvre and one for all the other manoeuvres; we have to solve
the problem of swit hing from one to another. Indeed swit hing between two

451
or more ontrollers in a dis ontinuous way ould ause instability of the non-
linear losed-loop system due to sharp transients and also, performan e may
deteriorate. Many papers have appeared on the problem but there is still no
general result.
In order to guarantee the ontinuity of the ontrol ommands, we introdu e
the following logi . The feedba k part of our ontroller is, as illustrated in
Se tion 28.2.2, a PI ontroller. Let us denote byKp1 and Ki1 the proportional
and integral matrix gains of the rst ontroller, and by Kp2 and Ki2 the pro-
~ the
portional and integral matrix gains of the se ond ontroller. Denoting by t
swit hing time instant between the two ontrollers, we have:

u(t~ ) = Ki1 xi (t~) + Kp1 x(t~) (28.10a)

+
u(t~ ) = Ki2 xi (t~) + Kp2 x(t~) : (28.10b)

If we add a feedforward input to the ontroller, say w , equations (28.10) be ome

u(t~ ) = Ki1 xi (t~) + Kp1 x(t~) + w(t~ ) (28.11a)

u(t~+ ) = Ki2 xi (t~) + Kp2 x(t~) + w(t~+ ) : (28.11b)

By setting

w(t~+ ) = (Ki1 Ki2 )xi (t~) + (Kp1 Kp2 )x(t~) + w(t~ )

the ontrol signal u turns out to be ontinuous.
Also the zero steady state error e = r y in the presen e of the additional
onstant input w is preserved by the presen e of the integrators in the loop.
Note that, in the digital implementation of the ontroller, unit delays (that
means additional states) would be used to store the values of the ontrol om-
mands during a sampling time.

28.2.4 The omputation of pw

Sin e lateral sti k dee tion demands velo ity ve tor roll rate whi h is not a
measured variable, we need to nd an e ient way to estimate or to ompute
this variable on the basis of the set of measured variables.
A nonlinear relation between the angular rates in the wind axes pw , qw and
rw and the angular rates in the body axes p, q and r as a fun tion of the angles
 and  is dened by:
0 1 0 10 1
pw os  os  sin  sin  os  p
 qw A =  os  sin  os  sin  sin  A  q A : (28.13)
rw sin  0 os  r
Sin e the  and  state variables are measured, we an use transforma-
tion (28.13) to obtain the a tual value of pw . Note that, when demanding
velo ity ve tor roll rate,  is kept almost onstant and  is minimised for
the manoeuvre to be satisfa torily performed. This means that the matrix in
transformation (28.13) turns out to be almost onstant; hen e the nonlinearity
introdu ed by this new sub omponent is expe ted to be su iently mild to not
deteriorate the losed-loop system performan e.

452
28.2.5 The nonlinear ompensation of the dynami pres-
sure variation
As usual when using a linear ontroller for a nonlinear plant, it doest not
guarantee the stability and performan e of the nonlinear losed-loop system;
indeed we are in the presen e of strong variations of the linearized models of
the plant around dierent operating onditions.
A possible way to avoid these problems is to ompensate some of the non-
linearities of the model to make as similar as possible, the linearized model
matri es in a large region of the operating envelope.
In the ase of the air raft it is possible to take into a ount the nonlinear
ee t of the velo ity and the altitude on the aerodynami for es and try to
ompensate it by means of a nonlinear s aling of some of the gains of the on-
troller. Indeed the aerodynami for es and moments generated by the ontrol
surfa es are linearly dependent on the dynami pressure q = 0:5V 2 .
If the LQ ontroller has been designed on the linearized model around one
operating point, where the nominal dynami pressure is q0 = 0:50V02 , when the
ontroller is applied in other operating onditions the ontrol law is amplied
or redu ed in relation to an in reasing or a de reasing dynami pressure.
In order to normalize the aerodynami surfa e ontrol power we s ale the
ontrol ommands at the output of the PI a tion with a term q0 =q.

28.2.6 Demand shaping lters

As usual in the aeronauti al appli ations, an important rle is played by the
DS (Demand Shaping) lters (see for example [28 and [3).
In our ontroller s heme, the lters have been used to a hieve the following
tasks:

i) to avoid the saturation of the ontrol ommands;

ii) to allow the satisfa tion of the Gibson riteria and the spe i ation on
dropba k a ording to the requirements given in Se tion 27.3.5.

Indeed, if the pilot inputs a sudden demand on one of the variables to be

ontrolled, it is possible that the ontrol ommand ould exhibit peaks or high
rate of variations. When dealing with a nonlinear plant, where saturations and
rate limiters are a ting, it may happen that the losed-loop system state is
brought out of the stability region.
On the other hand, the requirements on the input-output behaviour of the
plant expressed by Gibson riteria an be a omplished by introdu ing linear
lters at the input of the plant with an adequate pla ement of their poles and
zeros.
Two kind of lters have been onsidered in our ontroller:

1. a nonlinear lter whi h avoids the demand to be too high or too fast for
some riti al points of the ight envelope;

2. a linear lter onsisting of a stable linear dynami system, whose poles

and zeros lo ation allow for the satisfa tion of the Gibson riteria.

453
28.2.7 Swit hing logi on the  and Ma h values
The linear nature and the low order of the feedba k PI ontroller proposed in
Se tion 28.2.2, as said, has the drawba k that the gains designed around one
linearized model of the air raft are not assured to work well in the whole ight
envelope: therefore ontroller gain s heduling is needed. This is mainly due
to the fa t that, in the ase of high angle of atta k, the air raft has strong
oupling between lateral and longitudinal dynami s.
In prin iple we tried to s hedule only with Ma h number. If we refer to
straight and level ight onditions, the ight envelope is redu ed to be only
dependent on two variables, namely the Ma h number and altitude. Indeed, 
is for ed to be always zero while  is univo ally determined by the solution of
problem (28.5). In level ight,  is de reasing with the Ma h number, hen e
we have that at high Ma h number there are no ouplings between lateral and
longitudinal dynami s. Nevertheless it happens, during transients, that the
air raft is for ed to y at high angle of atta k even if the Ma h number is high.
In su h ases we have that the ontroller designed for straight and level ight,
due to the ouplings between longitudinal and lateral dynami s, does not work
adequately. Hen e it is ne essary to s hedule the ontroller gains also as a
fun tion of the angle .
In summary we have that a rst partition of the ight envelope is based on
the Ma h number. A se ond partition is based on the  values. The number
of PI gains to be stored in the FCC turns out to be 2nM n :
i) 2 for ea h point of the envelope based on the pilot demand (MOSL);

ii) nM is the number of subregions based on the Ma h number, (in our ase
nM = 2 sin e the envelope is divided into two subregions orresponding
to Ma h> 0:27 and Ma h0.27);

iii) n is the number of subregions based on the  value for ea h Ma h

number (in our ase n = 2 for  > 25 deg and   25 deg).

Assume now that we have found the optimal ontrollers to over the whole
ight envelope; in order to guarantee the ontinuity of the ontrol signal when
swit hing between two dierent sets of gains, on e again we use the auxiliary
variable w introdu ed in Se tion 28.2.3.

28.2.8 The observer

As spe ied in Se tion 27.3.3, an un ertain but onstant error on  and  in
the interval [ 2; +2 deg is present.
Sin e the availability of the omplete measurement of the state of the six
degrees of freedom (DOF) air raft model guarantees good performan e and
robustness properties of the losed-loop system, a reasonable way to operate is
that of identifying the measurement error value when ying straight and level
and then an elling it during the manoeuvre. To a hieve the dete tion of the
error we suggest to use a Linear Quadrati Observer (LQO).
The design of the LQO, as usual, results in hoi e of some weighting matri-
es. Our obje tive is to estimate the a tual values of  and  on the basis of

454
the available information about the ontrol inputs and the measured outputs
of the air raft.
Let us onsider the linear model of the air raft in the form 28.6, where the
states and the ontrol inputs are dened a ording to Se tion 28.2.1. The lassi-
al Luenberger observer is des ribed by the following linear equation (see [140):

x~_ = Ax~ + Bu u + L(y C x~) (28.14a)

y~ = C x~ (28.14b)

where x~ and y~ = ( ~ ~ )T are the estimated state and output ve tors and L
is the gain matrix of the observer. If we design the gain of the observer with
an LQG strategy we have to dene the weighting matri es Q~ and R~ in the
following Ri ati equation

A + AT + Q~ C T R~ 1 C  = 0 : (28.15)

Sin e this is a deterministi problem, the degree of freedom oming from the
hoi e of Q~ and R~ an be used to emphasize the estimate of  and  and to
speed up the onvergen e of the estimator.
Hen e, when the air raft is in a straight and level ight ondition, the LQO
onverges as soon as possible to the estimates ~ and ~ and the dete tion of the
errors is given by the dieren es ~ m and ~ m , m and m being the
measured variables. Until the next straight and level ight of the air raft, the
error is assumed to be known.
This kind of pro edure has three main drawba ks:

i) as said, the speed of onvergen e of the estimate has to be as fast as

possible; indeed it may happen that the air raft remains in a straight
and level ight ondition only for a few se onds;

ii) if, when performing the estimation, the measurement y is ae ted by
noise there ould be an in orre t estimation of the variables;

iii) the observer matri es annot be onsidered onstant over the whole op-
erating envelope be ause they are representative of the nonlinear plant
behaviour only in a neighbourhood of the design point.

To over ome problem i) we have to adequately hoose the weighting matri es

of the ontroller. Indeed if we try to speed up the time response of the observer
we ould have a de ay of the estimation performan e due to the fa t that the
measurement noise enters into the bandwidth of the ontroller. Hen e, as
always, it happens in the hoi e of the weighting matri es, a good trade-o has
to be found.
Con erning problem ii), we have veried via simulations that, if the distur-
ban es are zero mean, the designed estimator works well.
Finally to over ome problem iii) we have to s hedule the observer matri es
on the basis on the Ma h number and the altitude so as to over the whole
operating envelope.

455
28.3 The Translation of HIRM Design Criteria
into Method Dependent Obje tives
As previously said, one of the drawba ks of the LQ ontrol te hinque is that
it is not possible to take dire tly into a ount spe i ations on performan e
and robustness as spe ied in the HIRM problem denition Chapter 27. As
explained in Se tion 28.2.2, the designer has to work on the weighting matri es
to try to a hieve his obje tives; this trial and error pro edure to synthesize a
suitable ontroller ould take a very long time. In order to save part of this time
we introdu ed some auxiliary fun tions, dependent on the design parameters,
whi h allow us to he k, at ea h trial, if the given requirements are satised.
We dened two kinds of auxiliary fun tions:

 the fun tions denoted by gi translate the more stringent spe i ations of
the ontrol problem; these fun tions are positive if the given requirement
is not satised;

 the fun tions denoted by fi translate less stringent spe i ations; in par-
ti ular the loser fi is to zero the loser we are to the omplete satisfa tion
of the spe i ation.

Hen e the designer has to look, at ea h trial, at the gi and the fi fun -
tions to understand in whi h dire tion to move for further trials. Let Q^ =
^
diag(^q1 ; :::; q^n ) and R = diag(^r1 ; :::; r^m ) be the weighting matri es of the LQ
problem dened in Se tion 28.2.2, n and m being the number of states and
inputs of the augmented system (A; ^ B^ ); the design parameter ve tor is dened
as xopt = ( q
T
^1 ; :::; q^n ; r^1 ; :::; r^m ) . The following optimization problem has to
be solved

min
x
f (xopt ) s:t: (28.16a)
opt

gi (xopt ) < 0 ; i = 1; ::; n ; (28.16b)

where f (xopt ) =
P 2
i=1;::;no qi fi (xopt ) and gi (xopt ) are the fun tions to be min-
imized and the onstraints respe tively, qi are weighting elements, n is the
number of onstraints and no is the number of terms in the ost fun tion whi h
ome out from the performan e and robustness requirements. The solution of
problem 28.16 an also be automated by means of a numeri al optimization
pro edure.
In the following we will des ribe some of the fi and gi fun tions utilised to
translate the majority of the HIRM requirements.
Re all that the Gibson riteria, as well as the spe i ations on the  drop-
ba k given in Chapter 27, have been satised open-loop, by means of the de-
mand shaping lters. In parti ular, we will detail the te hnique used to hoose
the parameters of a rst order linear time-invariant lter for the q-demand
hannel.

Modelling Errors Considerations. Sin e the design is performed on the

linearized model of the plant, in the synthesis phase we onsidered the modelling

456
errors as un ertainties entering the system matri es of the air raft. Let us
onsider the un ertain linearized model of the air raft in luding sensors and
a tuators in the form

x_ = A(paer )x + Bu (paer )u (28.17)

y = Cx (28.18)

where paer is the ve tor of the model un ertainties whi h is assumend to belong
to the hyperretangle Paer a ording to Se tion 27.3.3.
We an he k the stability of the losed-loop system in presen e of the un-
ertainties paer 2 Paer by analysing the losed-loop eigenvalues for the 2
16
(i)
verti es of Paer , say paer ; sin e for many parameters the range of variation is
small, stability on the verti es should guarantee stability for all values of the
un ertainties. As the stability of the perturbed system is a stringent spe i a-
tion, it will be onsidered in an auxiliary fun tion whi h takes into a ount the
maximum real part of the eigenvalues of the losed loop system ACL(paer ):

gi (xopt ) = max(real(eig(ACL (p(aer

i) )))) ; i = 1; ::; 216 (28.19)

Stability Analysis under Perturbation at the Input of the Plant. The

stability test required in Se tion 27.3.4 deals with the simultaneous and inde-
pendent gain and phase osets at the input of ea h one of the 6 a tuators
as shown in Figure 27.6. In Chapter 4 it is re alled that the LQ ontroller
guarantees robust stability against purely real or purely imaginary matrix per-
turbations at the input of the plant while it does not guarantee stability for
mixed real- omplex perturbations, as those onsidered in the HIRM problem.
Nevertheless, to take into a ount the robustness requirements, further aux-
iliary fun tions have been introdu ed. Let us onsider the region plotted in
Figure 27.7 (Figure 27.8 in luding toleran es), say F  IR2 ; our matrix pertur-
bation is

P (pp ) = diag(K1 e j1 ; :::; K6 e j6 ) ( Ki i ) 2 F (28.20)

where

pp = ( K1 ; 1 ::: K6 ; 6 )T ; pp 2 F = F 6
an be seen as a ve tor of un ertain parameters.
Consider now the perturbed system matrix transfer fun tion Gp (s; pp ) =
G(s)P (pp ); an analysis of the stability of the perturbed system an be arried
out by onsidering it for the 2
12 verti es of the set F , say pp(i) , i = 1; : : : ; 212 .
h i
Let Ni
(i) be the Ni hols plot of det Gp (s; p(pi) )K (s) whi h, in a parametri-
zed form, is des ribed by the equations

Mag(i) = Mag(i) (!) (28.21)

P h(i) = P h(i) (!) : (28.22)

457
 We introdu e a new fun tion whi h represents the relative distan e in the
Ni hols plane between Ni (i) and the point ( 180; 0) in the interval of interest
! 2 [!1 ; !2 (see Figure 28.2):
8 q
< min! qMag(i) 2 + (P h(i) + 180)2 if Ni (i) passes under zero ;
dNi (i) =
:
+ min! Mag(i) 2 + (P h(i) + 180)2 if not :
(28.23)
If we have a plant with no poles with positive real part a ondition for
stability is that the 212 fun tions

g(216 +i) (xopt ) = dNi (i) i = 1; ::; 212 (28.24)

be negative.

Single Loop Analysis: Ni hols Plots. In Se tion 27.3.5 a performan e

spe i ation is given in terms of the single-input single-output Ni hols plot of
the frequen y response between ea h a tuator demand and the orresponding
error signal, obtained by breaking the loop at the point shown in Figure 27.6
while leaving the other loops losed.
These Ni hols plots should avoid the regions shown in Figure 27.9 (no tol-
eran es applied) and in Figure 27.10 (toleran es applied).
Let us onsider the parametrized equation of the Ni hols plot obtained
by breaking the i-th hannel of the loop in the frequen y interval of interest
! 2 [!1 ; !2 , say Ni 2(i)
Mag2(i) = Mag2(i)(!) (28.25)

P h2(i) = P h2(i) (!) (28.26)

and let (Mag (t); P h (t)) t 2 [0; 1 be the parametrized equation of the ontour
of the region to be avoided, denoted by D. We an establish a relative distan e
between the Ni hols plot and the region D (see Figure 28.3):
8 p
>
>
min!;t T(Mag Mag2(i))2 + (P h P h2(i) )2
(i) D = ; ; ! 2 [!1; !2 ; t 2 [0; 1 ;
<
if Ni 2p
d2(D; Ni 2(i) ) =
>
>
:
+ mint Mag Mag2(i)(~!))2 + (P h P h2(i))2 (~!)
(T
if Ni 2(i) D 6= ; ; t 2 [0; 1 ;
(28.27)
where !~ = (~!1 !~2 )=2, and !~1 and !~ 2  !~ 1 are the values of ! for whi h
Ni 2(i) interse ts the ontour of D.
We an now dene 12 new onstraints to be onsidered (6 without toleran es
and 6 with toleran es) whose negativeness guarantees the satisfa tion of the
performan e requirement about the Ni hols plots:

g(216 +212 +i) (xopt ) = d2(D; Ni (i) ) i = 1; ::; 12 : (28.28)

458
 40 10

30 8

6
20
d>0 4
10 d2>0
2
Mag (db)

Mag (db)
[180,0]
0 0
d2<0
2
10
d<0
4
20
6

30
8

40 10
300 250 200 150 100 50 200 190 180 170 160 150 140 130 120 110
phase (deg) phase (deg)

Figure 28.2: Distan e fun tion Figure 28.3: Distan e fun tion d2
dNi (i) from instability for performan e

Physi al Considerations. In Se tion 27.3.5 other requirements, whi h take

into a ount the physi al behaviour of the plant, are given. Due to the nonlinear
nature of these spe i ations it is not possible to take a ount any of them
dire tly in the LQ ontroller synthesis. Moreover, they an be onsidered in
the formulation of a number of auxiliary fun tions, part of them omputed
on the basis of numeri al simulations performed on the losed-loop nonlinear
system.
As an example, onsider the two spe i ations given on the sideslip time
response; the former on erns the oupling between roll and sideslip, while the
latter regards the response to a sideslip demand.
Let  (t; xo pt) be the sideslip response of the losed-loop nonlinear system to
a roll rate demand in the time interval [0; 10, obtained for the design parameter
xopt . We introdu e the following auxiliary fun tion

f5 (xopt ) = max  2 (t; xopt ) : (28.29)

t2[0;10

To satisfy the se ond requirement we make referen e to the region shown

in Figure 27.16, say B. The ontour of this region is dened by the following
parametri equations

L = L(t) t 2 [0; 10 (28.30)

U = U (t) ; (28.31)

whi h represent the lower and upper bound of B respe tively. Performing a ten
se ond simulation of the nonlinear losed-loop system, we obtain the sideslip
time response  (t; xopt ). Let us dene a time dependent distan e fun tion in
the form:
8
< 0 if L(t)   (t; xopt )  U (t) ;
d3(t; xopt ) = L(t)  (t; xopt ) if  (t; xopt )  L(t) ; (28.32)
:
 (t; xopt ) U (t) if  (t; xopt )  U (t)

459
We introdu e the following auxiliary fun tion

f6 (xopt ) = max d3(t; xopt ) : (28.33)

t2[0;10

The Gibson Criteria. For the q demand- hannel, two spe i ations have
to be satised: the ratio of dropba k to steady-state pit h rate should be in the
interval [0; 0:25 (see Se tion 27.3.5 and Figure 27.15) and the Gibson plots spe -
ied in Se tion 27.4.2 should pass through the regions shown in Figures 27.13
and 27.14 of the HIRM manual. The rst one is a time domain spe i ation,
while the se ond one is a frequen y domain spe i ation. We propose to op-
erate in the time domain (whi h is oherent with our whole design te hnique)
and to verify a posteriori the satisfa tion of the Gibson plots requirements.
Let us onsider the transfer fun tion of the linearized HIRM system on
the q demand- q hannel, namely Wq (s). The relation between this transfer
fun tion and the transfer fun tion orresponding to the hannel q demand-
isW (s) = Wq (s)=s. Hen e, for ing W with a unit step signal is equivalent
to for e Wq (s) with a ramp signal of unit slope. From linear system theory
we know that if an asymptoti ally stable system W (s) is for ed with a ramp
signal, the asymptoti behaviour of the output is generally (if no zeros in the
origin are present) still a ramp, having the following equation


d 
yr (t) = W (s)js=0 t + W (s) : (28.34)
ds s=0
From the above onsideration and from Figures 27.14 and 27.15 it is lear that
the transfer fun tion Wq has to satisfy the following requirements

Wq (0) = 1 (28.35a)

d 
0  Wq (s)  0:25 : (28.35b)
ds s=0
In order to a hieve this obje tive by means of a q demand shaping lter
we propose to nd, by means of the numeri al solution of an optimization
problem, a rst order, asymptoti ally stable, unit stati gain, transfer fun tion
in the form
ps+z
Wf (s; z; p) =
zs+p
whi h satises the ondition

Wqf (0; z; p) = 1 (28.36a)


d 
0  Wqf (s; z; p)  0:25 (28.36b)
ds s=0
where z and p are parameters to be optimized and
Wqf (s; z; p) = Wf (s; z; p)Wq (s) :

460
28.4 The Des ription of the Design Cy le
Two points are ru ial in the design pro ess: the s heduled nature of the on-
troller whi h requires a denition of the s heduling regions, and the di ulty
of onsidering, during the LQ based matrix gains design, the whole set of spe -
i ations given in Chapter 27.
The design y le an be synthesized by means of the following pro edure

1. Choose a re tangular region, say Ej , in the straight and level ight enve-
lope (Ma h-altitude plane) to design a ontroller with xed (non s hed-
uled) parameters;

2. Design a xed set of matrix ontroller gains via the trial and error pro-
edure des ribed in Se tion 28.3 to take into a ount the HIRM require-
ments;

3. Design, if ne essary, dierent gains of the ontroller depending on the

angle of atta k;

4. Test the performan e of the ontroller designed by means of the evaluation

fun tions fi and gi introdu ed in Se tion 28.3. If they are not satisfa tory,
redu e the region Ej and go to point 1.;
5. Sele t a new region adja ent to (or in luding) the previous one and repeat
points 1. to 4. until the whole igth envelope is overed;

6. Design the demand shaping lters looking at the Gibson plots of the
losed loop system;

7. Design the LQO for the error dete tion for a su iently large number of
points within the operating envelope;

8. Test the performan e of the omplete losed-loop nonlinear system by

means of numeri al simulation.

The numeri al tools whi h were used to support the above design y le
are well-established for solving the LQ problems, with some numeri al opti-
mization odes, problem oriented, developed to qui ken ea h step of the design
pro edure. Among these, a Simulink based User Interfa e for the Design and
Analysis of the Flight Control System has been built, by whi h the designer
an run MATLAB fun tions developed for the HIRM problem.

28.5 Numeri al Results

In this se tion the set of gures, gure 28.4 to gure 28.9, show the results of the
Automated Evaluation Pro edure that reveal the performan e and robustness
of the proposed ontrol s heme.

461
 Parametric uncertainty, Nichols plot, Mach 0.4, 10000 ft Parametric uncertainty, Nichols plot, Mach 0.24, 20000 ft
20 20

15 dts 15 dts
. dtd . dtd
10 10
o dcs o dcs
OpenLoop Gain (dB)

x dcd

OpenLoop Gain (dB)

5 x dcd
5
+ dr + dr
0 * thr 0 * thr

5 5

10 10

15 15

20 20
350 300 250 200 150 100 50 0 350 300 250 200 150 100 50 0
OpenLoop Phase (deg) OpenLoop Phase (deg)

Figure 28.4: Parametri un er- Figure 28.5: Parametri un er-

tainty, Ni hols plot, Ma h 0.4, al- tainty, Ni hols plot, Ma h 0.24, al-
titude 10000 ft titude 20000 ft

28.6 Con lusions

To evaluate the suitability of the proposed ontrol s heme and of the related
design y le for problems su h as the HIRM design ontrol problem, we would
like to summarize the aspe ts whi h, in our opinion, an be onsidered to be
the main advantages and the main drawba ks of the method.
Of ourse, we have the following advantages:

i) the ontroller has a low order and the s heme is quite simple to be imple-
mented, sin e it is essentially a PI ontroller; this kind of stru ture also
simplies the understanding of its physi al a tion on the air raft;

ii) it does not require a deep theoreti al ba kground to apply the proposed
method; furthermore the onditions that the plant should verify are very
minor and an be easily he ked;

iii) the used software is well established (all the software developed for the
LQ methods) or quite easy to implement;

iv) most of the design requirements are easily satised; some other require-
ments an be satised by tuning the ontroller parameters;

v) re-design of ight ontrol laws with this method should be straightfor-

ward, provided that the physi al problem is not signi antly hanged.

On the other hand, the main drawba ks are:

i) the design is partially based on a trial and error pro edure whi h does
not guarantee a short design time; however, based on our experien e, the
pro edure does not need many iterations;

ii) some of the given requirements annot be easily taken into a ount in the
design phase;

462
 1) HIRM problem understanding 12%
2) Study of the literature on the LQ methods 5%
3) Translation of HIRM requirements into LQ spe i ations 10%
4) Denition of the ontroller stru ture 12%
5) Tuning of the ontroller parameters 12%
6) Software development 16%
7) Report writing 18%
8) Running the evaluation pro edure 10%
9) Others 5%
Total 100%

Table 28.1: Time spent for the HIRM design

iii) a omplete simulation ampaign on the nonlinear losed loop system is

needed to validate the ontroller at ea h ight ondition, but it is ertainly
true that this last point is also ommon to any ight ontrol system design
(as stressed for example in [3), espe ially if linear plant oriented.

We would nally like to give some information on erning the design time
spent during this a tivity by our proje t team. The total amount of work on
the design problem was about 45 man/weeks. In Table 28.1 the time spent for
ea h developed a tivity is shown.
From this table it is evident that the majority part of the design time has
been spent to understand the HIRM problem, hoose the ontrol s heme and
translate the whole set of requirements into method dependent fun tions. On e
the ar hite ture of the ontroller was hoosen, only 12% of the time was spent
tuning the design parameters.

463
 pw q az
0.05 8 8

6 10
4
12

(m/s^2)
(deg/s)

(deg/s)
0 2
14
0

2 16

0.05 4 18
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

Va alpha x 10
3 beta
101.5 22 8

101 20 6

100.5 18 4

(deg)

(deg)
(m/s)

100 16 2

99.5 14 0

99 12 2
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

dts dtd dcs

5 0.03 0

10 0.02 5
(deg)

(deg)

(deg)
15 0.01 10

20 0 15

25 0.01 20
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

x 10
3 dcd dr thr
2 0.06 60

0 50
0.04
2 40
(deg)

(deg)

(%)
4 0.02 30

6 20
0
8 10

10 0.02 0
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

Assessment manoeuvre: pitch rate demand, Mach 0.3, 5000 ft

Figure 28.6: Assessment manoeuvres: pit h rate demand at M=0.3, h=5000

feet

pw q az
80 10 5

60
0
5
40
(m/s^2)
(deg/s)

(deg/s)

5
20
0
10
0

20 5 15
0 2 4 6 0 2 4 6 0 2 4 6
time (s) time (s) time (s)

Va alpha beta
104 15 1

103 0
10
102 1
(deg)

(deg)
(m/s)

101 2
5
100 3

99 0 4
0 2 4 6 0 2 4 6 0 2 4 6
time (s) time (s) time (s)

dts dtd dcs

10 5 10

5 0
5
5
0
(deg)

(deg)

(deg)

10 0
5
15
5
10 20

15 25 10
0 2 4 6 0 2 4 6 0 2 4 6
time (s) time (s) time (s)

dcd dr thr
15 10 4

10 5 3
(deg)

(deg)

(%)

5 0 2

0 5 1

5 10 0
0 2 4 6 0 2 4 6 0 2 4 6
time (s) time (s) time (s)

Assessment manoeuvre: roll rate demand, Mach 0.3, 5000 ft

Figure 28.7: Assessment manoeuvres: roll rate demand at M=0.3, h=5000 feet

464
 pw q az
15 0.2 9

10
0.1 9.5
5

(m/s^2)
(deg/s)

(deg/s)
0 0 10

5
0.1 10.5
10

15 0.2 11
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

Va alpha beta
100.8 14.5 15

100.6
14 10

100.4

(deg)

(deg)
(m/s)

13.5 5
100.2

13 0
100

99.8 12.5 5
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

dts dtd dcs

11.8 5 6

12 0
6.5

12.2 5
(deg)

(deg)

(deg)
7
12.4 10

7.5
12.6 15

12.8 20 8
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

dcd dr thr
1 20 12

15 10

8
0.5 10
(deg)

(deg)

(%)
6
5
4

0 0 2

5 0
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

Assessment manoeuvre: sideslip demand, Mach 0.3, 5000 ft

Figure 28.8: Assessment manoeuvres: sideslip demand at M=0.3, h=5000 feet

pw q az
0.01 8.5

0 9
0.01 0
9.5
(m/s^2)
(deg/s)

(deg/s)

0.02
10
0.03
0.5
0.04 10.5

0.05 11
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
time (s) time (s) time (s)

Va alpha x 10
3 beta
160 14 8

150 6
12
140
4
(deg)

(deg)
(m/s)

130 10
2
120
8
110 0

100 6 2
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
time (s) time (s) time (s)

dts dtd dcs

9 0.01 6

7
10 0.005

8
(deg)

(deg)

(deg)

11 0
9

12 0.005
10

13 0.01 11
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
time (s) time (s) time (s)

x 10
3 dcd dr thr
6.5 0.06 100

80
7 0.04

60
(deg)

(deg)

(%)

7.5 0.02
40

8 0
20

8.5 0.02 0
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
time (s) time (s) time (s)

Assessment manoeuvre: air speed demand, Mach 0.3, 5000 ft

Figure 28.9: Assessment manoeuvres: velo ity demand at M=0.3, h=5000 feet

465
29. The H1 Loop-Shaping Approa h
George Papageorgiou1 , Keith Glover 1

and Ri k A. Hyde 2

Abstra t. The H1 loop-shaping design pro edure (referred to

hereafter as LSDP) was used to design xed gain ontrollers for
the HIRM. It was found that a single xed gain ontroller designed
for the linearisation ini3005, performed adequately over the whole
ight envelope of the HIRM. As s heduling would most probably
be required for a real ghter air raft with a mu h wider ight enve-
lope, the ontrol law has been developed with a stru ture suitable
for s heduling. As the longitudinal and lateral dynami s are es-
sentially de oupled, two separate ontrollers were designed. The
lateral ontroller has an inner loop for stabilisation and an outer
loop for tra king. The longitudinal ontroller has one single loop
that provides both fun tions. A se ond xed gain ontroller was
designed for the HIRM using the multi-model approa h des ribed
on pp. 67-72 in [247. This method enables the designer to nd a
xed gain ontroller that makes the normalised oprime fa tor ost
fun tion at over the whole ight envelope. Details about how to
apply this method an be found in [191. The design team found it
relatively easy, even with the limited time available, to satisfy the
spe i ations set out in Chapter 27.

29.1 Introdu tion

The LSDP is an intuitive method for designing xed gain robust ontrollers. A
tutorial on H1 loop-shaping is given in Chapter 7. A ontroller that has been
designed using H1 loop-shaping provides robust stability to oprime fa tor
un ertainty. Coprime fa tor un ertainty is a general type of un ertainty mu h
in the same way that single-input single-output (SISO) gain and phase margins
are. Therefore, when there is little detailed knowledge about the un ertainty
present in a plant the LSDP is a good method for designing robust ontrollers.
The dieren e to gain and phase margins is that oprime fa tor un ertainty
an be used to dire tly address robustness in systems with multiple feedba k
loops, i.e. multivariable systems (see pp. 240-244 in [266).
The LSDP has been used in a variety of appli ations and studies. Most
relevant to the HIRM design hallenge is the work in [120 where a ight on-
1
Cambridge University Engineering Department, Cambridge CB2 1PZ, England
2
Cambridge Control Limited, Cambridge CB4 4WZ, England

466
trol law was developed for the DRA Bedford Resear h Harrier XW175 and
subsequently ight tested. The ontrol law performed well and is the subje t
of on-going work. The experien e of developing this ontrol law showed that
H1 loop-shaping has a number of key attributes whi h make it parti ularly
suitable for this type of appli ation. Perhaps the most important attribute
is that the resulting ontroller an be written as an exa t plant observer plus
state feedba k. This stru ture allows gain s heduling, as designs at dierent
operating points have the same state-spa e stru ture and hen e the observer
and state feedba k gains an be linearly interpolated. The observer also allows
the handling of input limitations (e.g. authority and rate limits) by driving
the observer with a hieved plant inputs rather than demanded ones. A third
very important attribute of H1 loop-shaping ontrollers is that the observer
stru ture adds visibility, in that physi al units and interpretation an be ap-
plied to the ontroller oe ients and states. This may be advantageous with
respe t to ight learan e. The advantages of the observer stru ture are fully
exploited for the HIRM ontrol law design.
H1 loop-shaping has mu h in ommon with the design approa h urrently
used by industry, sometimes referred to as  lassi al ontrol. The loop-shaping
part of the design pro edure is arried out in exa tly the same way that las-
si al design is arried out: inputs and outputs are mat hed up, and single
loop shaping is arried out to ensure that low frequen y gain is large enough,
roll-o at ross-over is not ex essive and that su ient high frequen y roll-o
is provided. On e this is done, then the optimal H1 loop-shaping ontroller
is synthesised for this so- alled weighted plant, see Se tion 7.2. Hen e, it is
perfe tly possible to take a lassi al design, and to augment it with a orre-
sponding H1 ontroller whi h will then modify the feedba k stru ture so as to
allow for the multivariable nature of the system.
When designing using lassi al ontrol where the system has inherent ross-
oupling (as for example with most yaw - roll augmentation systems), orre-
sponding ross-terms are put into the ontroller. Design of these terms is not
always straightforward in that their ee t on the feedba k loops is not ad-
dressed dire tly, and some iteration may be required. With the LSDP, these
ross-terms an be left to the H1 synthesis part of the design. In the authors'
view, one of the prime motivations for H1 loop-shaping is the potential for de-
sign time redu tion, parti ularly when dealing with multi-input multi-output
systems with strong ross- oupling.

29.2 The Controller Ar hite ture

The rst stage of the design pro ess is to sele t the ontrol law ar hite ture.
Sele tion of the ar hite ture is an essentially design method independent task,
and the reasoning used for the presented design is mu h the same as would be
used for a lassi al ontrol law. However, H1 loop-shaping and related robust
optimal ontrol methods ould have been used to help sele t the ar hite ture by
looking at the robustness impli ations of ea h andidate ar hite ture, e.g. ex-

467
amination of the robustness impli ations of dierent anard and taileron blend-
ing s hemes would be possible. Limited time available pre luded this type of
analysis. If a full ontrol law design for a prototype or produ tion air raft was
undertaken, ontrol law stru ture sele tion with referen e to robustness impli-
ations is denitely re ommended. In sele ting the stru ture here, robustness
requirements are taken into a ount in a more heuristi way from knowledge
of the system to be ontrolled.
All of the a tuators available for the design hallenge are used with ex-
eption of the dierential anards. This is be ause omitting any of them will
ne essarily ompromise performan e in terms of a hievable for es and moments.
The reasons for not using the dierential anards are given in Se tion 29.2.2.
There are four primary feedba k loops to design, three rotational ones and
airspeed. Multivariable ontrol allows the designer to design all four simulta-
neously. However, the longitudinal motion  pit h and airspeed  is essentially
de oupled from the lateral loops  yaw and roll. What oupling there is be-
tween lateral and longitudinal motion is due to kinemati ross- oupling and/or
asymmetri aerodynami for es due to, for example, dierent ow regimes over
ea h of the wings. The linearisations provided for wings level, steady ight do
not apture these ee ts. Benets from designing on the omplete 4-input 4-
output system are likely to be more prominent in a Linear Parameter Varying
(LPV) framework [259 within whi h these oupling terms ould be modelled.
For example, parametri dependen e on roll rate ould be modelled and hen e
designed for. However, given the time onstraints, an LPV solution was not
investigated and hen e the de ision was taken to separate the longitudinal and
lateral ontrol law designs.
Figure 29.1 shows the top level SIMULINK spe i ation of the ontroller.
The two H1 loop-shaping ontrollers are ontained within the lateral and lon-
gitudinal H-infinity ontroller blo ks. They are implemented in dis rete
time observer form and hen e have two sets of inputs, the measurements and
the a hieved air raft inputs. The pre- ompensator weights, longitudinal W1
and lateral W1, ontain all the integrators, phase advan e terms and roll-o
terms designed in the same way as for a lassi al ontrol law. They are im-
plemented in a modied Hanus self- onditioned form (see Chapter 7 in [120
and [110). This is exa tly the stru ture used for the Harrier ontrol law de-
veloped in [120. Note the two s aling blo ks in the feedba k paths. These are
used to trade-o the relative amounts of oupling whi h are to be tolerated
between outputs, e.g. s aling speed in knots and pit h rate in degrees implies
that a 1 knot variation in airspeed is as equally undesirable as 1 /s oupling
in pit h rate. The outputs blo k implements rst order high frequen y roll-o
lters on the p, q and r measurements. The ut-o frequen y is 50 rad/s.

29.2.1 Longitudinal Controller

The primary feedba k variables used to design the longitudinal ontroller are
pit h rate q and airspeed V. This hoi e is straightforward in that these are
the quantities the pilot wishes to ontrol. Use of pit h attitude  to stabilise

468
 incidence limits normal acceleration
limits
5
RH inceptor +
K
(stick demand
in Newtons) longitudinal Hinfinity
long scaling
controller
2 Mux
4
Longitudinal
RH inceptor + LH inceptor
(pitch rate demand + (speed demand in m/s)
in deg/s)

longitudinal dc gain
compensation longitudinal
W1
7
command filter
6 K Measurements Mux
1
Lateral stick force to
RH inceptor Control
deg/s demand outputs [1/hrv2
(roll stick force)
1/rho
hrv2
+ rho]
1 +
actuator
Lateral + demands
RH inceptor +
(roll rate demand
in deg/s) outputs
convert to lateral dc gain
velocity vector compensation
roll rate lateral
demand W1
3
pedals
(sideslip
demand in
degrees) sideslip
K
controller
lateral Hinfinity
lat scaling
controller

Figure 29.1: The ontrol law ar hite ture

the air raft would require phase advan e, whi h in ee t dierentiates the mea-
surement over some frequen y range. This would produ e a noisier signal than
the measured q and a less robust design. Furthermore, the dynami s of the
pit h attitude sensor are slower than those of the q sensor and pit h attitude
an not be used at large roll angles.
The HIRM has tailerons and anards available for longitudinal ontrol.
There are several dierent strategies whi h ould be used to determine how
to apportion a required pit hing moment between the surfa es. Before sele t-
ing a strategy a number of onsiderations must be taken into a ount:

 The anards have mu h faster dynami s, and hen e an be used to higher

frequen ies with less phase lag.

 The tailerons an generate mu h larger pit hing moments.

 The tailerons generate a non-minimum phase ight path response, whereas

the anards produ e a response in the ommanded dire tion immediately.

 Produ ing ountering pit hing moments simultaneously from tailerons

and anards is ine ient.

 The ar hite ture may have stru tural loading impli ations.

 Consideration must be made of what happens when the surfa es rate or

authority limit. Rate limiting of the surfa es an lead to pilot-indu ed
os illations.

469
Two possible s hemes are:

 Frequen y blending of the two a tuators. Using a omplementary lter,

the higher frequen y omponent of the demand is fed to the anards, and
the lower frequen y omponent to the tailerons. In this way the anards
are used to obtain a fast initial response. The anard pit hing moment
is then transferred to the tailerons whi h take the low frequen y part of
the demand. In this way the trim is taken by the tailerons whi h have
the moment generating power. This is energy e ient in that the trim is
taken only on one surfa e and hen e the surfa es do not oppose ea h other.
Using the anards at higher frequen ies may also allow more bandwidth
to be extra ted from the system. Rate limiting an be addressed by ross-
feeding the demand not generated by the surfa e that has rate limited to
the other surfa e.

 Driving both surfa es in tandem. The inputs an be s aled su h that the

demand is a per entage of total travel so that both surfa es saturate at the
same point. One of the motivations for this approa h is that rate limiting
is less likely to o ur. If both surfa es ee t the demanded pit hing
moment they both have less far to travel than if one surfa e had been
used. A disadvantage is that the extra agility of the anards is not being
exploited. Additionally, small high frequen y disturban es drive the mu h
heavier tailerons whi h may be less energy e ient in terms of required
hydrauli power. A se ond advantage of this s heme is that a failure of
one of the surfa es still gives a system whi h generates pit hing moments
a ross the required frequen y range, albeit with redu ed authority.

Either of the two above strategies ould be employed. The rst s heme was
hosen for this design. The omplementary lter is of the form

s !f
F anard (s) = ; Ftaileron (s) = :
s + !f s + !f
These transfer fun tions are implemented in dis rete time. The anard demand
is also normalised with the gain N . This gain is su h that the gain per unit
demand to pit h rate at open-loop ross-over frequen y is the same for both
anard and taileron. Therefore if one surfa e saturates or rate limits, its de-
mand an be fed dire tly a ross to the other surfa e. All of the limiting and
ross-feeding o urs in the a tuator demands blo k, Figure 29.1.
A pit h attitude hold is implemented in the outputs blo k in Figure 29.1.
Although not listed as a spe i ation of the ontrol law in Se tion 27.3.2, some
kind of hold is required in pra ti e. This enables the pilot to go sti k-free if he
so wishes. Figure 29.2 shows the pit h attitude hold blo k, the output of
whi h is a pit h rate demand. In ee t the variable out_1 be omes q +0:25err ,
where err =  urr prev . The robustness properties of the losed-loop are not
altered signi antly by feeding ba k q + 0:25. This is be ause the  portion of
the signal does not modify the loop gain at ross-over too mu h. The attitude
hold is only engaged when the ag input, hold flag in Figure 29.2, is set to

470
 1
1/z
hold flag
last out
*
4 Abs < AND
phi
NOT
+
0.1 +
0. +
+
+ 1
0.1 radians 0.25
out_1
*
3
2
theta
q

Figure 29.2: The pit h attitude hold blo k

1 and the roll angle  is smaller than 0.1 rad. A pit h attitude hold is not
desirable at large bank angles.
The onstru tion of the total pit h rate demand, ommand filter blo k in
Figure 29.1, onsists of the following terms:

 Longitudinal sti k demand. The sti k ommands normal a eleration in

g. This is onverted into an equivalent pit h rate demand. The demand
is also fed through a pre- ompensator blo k whi h is used to meet the
pit h drop-ba k requirement, see Figure 27.15.

 Normal a eleration limiting term. If the normal a eleration limits are

ex eeded, the amount by whi h the limit is ex eeded is onverted into the
required pit h rate demand to remove the ex ess.

 In iden e limiting term. The ex ess in iden e is turned into a pit h rate
demand that will restore in iden e to within the spe ied limits.

Theattitude-hold engaged flag blo k is shown in Figure 29.3. This

blo k is ontained within the ommand filter blo k mentioned above. The
output of this blo k engage flag, is the input hold flag of the blo k depi ted
in Figure 29.2. The attitude hold is only engaged if the pit h rate is less than
1 /s, and the pit h rate demand is less than 0:1 /s (ee tively zero). On e
engaged, it stays engaged whatever the pit h rate, until the pilot ommands a
non-zero pit h rate.

29.2.2 Lateral Controller

The primary feedba k variables used to design the lateral ontroller are roll
rate p and yaw rate r. Roll angle and heading angle were not used as primary
feedba k variables for the same reason that pit h attitude was not used when
designing the longitudinal ontroller (see Se tion 29.2.1). In addition, note that
p and r are valid for any orientation in spa e, whi h is not the ase for Euler
angles. A velo ity ve tor roll rate and side-slip demand system is spe ied
in Se tion 27.3.2. The designer might therefore be tempted to use side-slip
as a primary feedba k variable. This will almost ertainly produ e poorer

471
 1 Abs
<
q_dem
pilot demand
zero
0.0017
0.0017

2 Abs
< OR AND 1/z 1/z 1
q pitch rate alphahold delay engage
less than q zero or engaged flag
1 degree/s alphahold
0.017 engaged
0.017

Figure 29.3: engage flag logi

1 0.
* 1
lateral stick degrees to radians p demand
demand
*
+
2 + 2
sideslip controller r demand
demand
3 f(u)
alpha sin

f(u)
cos

Figure 29.4: Lateral demands

performan e in that the side-slip measurement is both more noisy and a slower
measurement than r. These two ee ts will mean that less bandwidth would
be extra ted from the yaw loop. Designing a tight primary feedba k ontroller
using p and r robustly stabilises the air raft and provides an inner losed-loop
system around whi h an outer loop side-slip tra king system an be built. The
side-slip ontroller an be seen in the lower left-hand orner of Figure 29.1.
Its output is a yaw rate demand whi h enters the inner loop as illustrated in
Figure 29.4.
A yaw rate demand is also ross-fed from lateral sti k to ee t a velo ity
ve tor roll  the required term is sin  times the lateral sti k demand as is
illustrated in Figure 29.4. Omitting this ross-term would leave it to the side-
slip ontroller to reje t the side-slip indu ed when a roll rate is ommanded. As
the side-slip outer loop has a lower bandwidth than the inner loop, ex essive
side-slip oupling will o ur. The ross-term puts in a fast yaw rate demand
to a hieve the velo ity ve tor roll.
The HIRM has dierential anards and tailerons available for roll ontrol.
However, the anards are very inee tive in roll as they generate smaller for es,
and are lo ated mu h loser to the entreline of the air raft. Hen e using
anards requires large surfa e dee tions and gives little benet. Furthermore,
it limits their availability for longitudinal ontrol for whi h they have denite

472
benets over the tailerons. By studying the aerodynami s of the HIRM it an
be dedu ed that the dierential anards also have a very signi ant inuen e
on the ee tiveness of the symmetri al anards and tailerons thus reating a
robustness issue. Hen e, only the dierential tailerons are used for roll ontrol.
For yaw ontrol, only the rudder is available.

29.3 Controller Design

This hapter dis usses how the weighting fun tions and ontroller parameters
were sele ted to meet the design requirements.

29.3.1 Lateral Primary Feedba k Controller

In designing the feedba k ontroller the main obje tive is to push the ross-over
frequen y as high as possible, whilst retaining an a eptable level of robustness.
No referen e is made to the handling qualities required. This is be ause if the
feedba k ontroller is as fast as possible given robustness onstraints, design of a
pre-lter to meet handling qualities should be straightforward. A fast feedba k
ontroller redu es losed-loop response lag, and gives a robust response i.e.
one whi h should not vary signi antly with envelope operating point. If this
approa h gives better handling qualities than spe ied, the high bandwidth is
still justied in terms of the a hieved disturban e reje tion.
The argument against in orporating handling quality requirements when
designing the feedba k ontroller is that ne essarily a trade-o between robust
stability and handling qualities will be ee ted. One ould argue that the
primary purpose of the feedba k ontroller is to maximise robustness for the
spe ied open-loop ross-over frequen ies, i.e. disturban e reje tion properties.
The design pro edure an be divided into the following steps.

1. S ale the dierential taileron and rudder inputs by 1= 21 V 2 . This nor-

malises the moment generated so as not to vary signi antly with ight
ondition.

2. Sele t the lateral states from the linearisation i.e. v, p, r and . Append
this linearisation with the a tuator models, full order sensor models, anti-
aliasing lters and omputational delay. The full order sensor models are
used sin e the nal ontroller is model redu ed anyway.

3. S ale the outputs to ree t the oupling requirements. The s aling used

is ( -tools and MATLAB ommands are used)

>> out_s aling = diag([0.3 1);

A ouple of design iterations were arried out before arriving at this

s aling. The s aling ree ts the inherent oupling within the plant i.e.
that a unit oupling in roll happens a lot more easily than a unit oupling
it yaw. Attempting to s ale the outputs to make the oupling into yaw

473
 and roll approximately equal results in poorer robustness. In general,
trying to hange the dire tionality of the plant is not good pra ti e. The
roll rate p is augmented with the roll angle to give the output variable
p +  . This boosts the low frequen y gain, and enables a roll angle
hold to be ee ted for zero lateral sti k demand. The  term is removed
during a roll-rate demand. This an be justied (in terms of robust
stability) provided that  is hosen su h that the open-loop ross-over
is entirely set by the p part of the onstru ted output variable.
4. The desired ross-over frequen y for both loops is 10 rad/s. This is the
highest the ross-over frequen y an go before robustness margins are
ne essarily redu ed, due primarily to a tuator roll-o. To verify this,
a few design iterations at slightly higher ross-over frequen ies an be
arried out, and the resulting a hieved robustness margin , monitored.
Both loops have suitable roll-o rates at ross-over, and so all that is
required is to boost low frequen y gain, and add high frequen y roll-o
lters. The sele tion of the appropriate transfer fun tions is exa tly as for
50
a lassi al design. Both loops are rolled o with the lter
s+50 . The lters
are dis retised using a bilinear transformation with frequen y warping to
mat h the lters at 10 rad/s.

Low frequen y gain is boosted with the pre- ompensator

>> W_p = nd2sys([1 1,[1 0);

>> W_r = nd2sys([1 2,[1 0);
>> W_1 = daug(W_p,W_r);

6
10

5
10

4
10

3
10
singular values in db

2
10

1
10

0
10

1
10

2
10

3
10 2 1 0 1 2
10 10 10 10 10
log frequency

Figure 29.5: Singular values of the weighted plant

5. Multiply the two plant inputs with gains kw4 and kw6, an input s aling
whi h gives the required ross-over frequen ies. Plot the singular values
of the shaped 2-input 2-output plant, Figure 29.5. Che k that the desired
loop shapes have been a hieved.

6. Design the H1 loop-shaping ontroller, K1 , for the shaped plant. Che k

that the resulting robustness margin is su ient (typi ally  > 0:3 indi-

474
 ates a robust design). Che k the step responses. These are shown in
Figures 29.6 and 29.7.

Step on p Step on r
1.2 1.2

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0

0.2 0.2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
time (s) time (s)

Figure 29.6: Step on p demand Figure 29.7: Step on r demand

If gain s heduling is to be arried out, then the above pro edure is applied
to ea h of the design points. Typi ally, the dynami weighting fun tions will
not need to be hanged between design points, and so all that hanges are kw4,
kw6 and the ontroller K1 . The ontroller is then implemented in observer
form, and the state-spa e matri es gain s heduled using linear interpolation.
Details on how to do this an be found in [120.
For the HIRM design hallenge, the ight envelope is not that wide and
hen e a single xed gain ontroller an be used. The xed gain ontroller
used was the one designed for the operating point ini3005. This linearisation
orresponds to 12 in iden e, and is somewhere in the middle of the ight
envelope. The orresponding values of kw4 and kw6 for this ight ondition
were then used for all ight ases. The a hieved  was 0:32. Should the level of
robustness have proven insu ient, then the loop-shaping exer ise above ould
have been re-run with less stringent performan e requirements.

29.3.2 Side-slip Controller

An outer loop side-slip ontroller was designed using loop-shaping and nor-
malised oprime fa tor robust stabilisation. By examining the Bode plot of
yaw rate demand to sideslip, the following ompensator was determined

4(s + 2)
W1 =
s(0:1s + 1)
This has a rst order roll-o at 10 rad/s to attenuate side-slip sensor noise, and
additional low frequen y gain below 2.0 rad/s. The Bode plot of the shaped
loop is shown in Figure 29.8.
The resulting H1 ontroller was model redu ed to 4 states by arrying out
a least squares mat hing of the gain and phase plots of the full order ontroller.
This an be done using the MATLAB fun tion invfreqz.m. The a hieved
robustness margin is  = 0:42. This guarantees a gain margin of at least 2:4,

475
 5 1.2
10

4
10
1
3
10
0.8
2
10
singular values

1 0.6
10

0
10 0.4

1
10
0.2
2
10

3 0
10

4
10 2 1 0 1 2 0.2
10 10 10 10 10 0 1 2 3 4 5 6
frequency (radians/second) time (s)

Figure 29.8: Shaped loop Figure 29.9: Step on  response

and a phase margin of at least 45 (see Se tion 7.3 for the relevant formulae).
Figure 29.9 shows the side-slip step response whi h meets the spe i ation set
out in Figure 27.16.

29.3.3 Longitudinal Primary Feedba k Controller

The longitudinal ontroller initially had a multivariable stru ture with feedba k
variables q and speed. However, the dieren e in the required ross-overs for
these two loops (approximately a fa tor of 10) meant that there was little or
no advantage to having a multivariable stru ture. Hen e, two SISO ontrollers
were designed given that the added omplexity of the ross-terms in the MIMO
ase ould not be justied in terms of robustness or performan e. The q feed-
ba k was designed using the LSDP, and the velo ity feedba k using lassi al
loop-shaping. The losed-loop time response shows a good de oupled response
with no overshoot. Full details an be found in [191.

29.3.4 Pilot q - ommand Filter Design

The losed-loop step response gives little or no overshoot. In order to a hieve
the drop-ba k requirement, a pilot ommand lter of the following form was
used:
T s+1
ppilot (s) = 1
T2 s + 1
By putting T1 > T2 pit h rate overshoot is introdu ed. The values of T1 and
T2 were determined by performing a sear h over a range of values of T1 and T2 ,
and then sele ting all pairs whi h give a drop-ba k of approximately 0:125 s.
The other handling quality riteria were then examined for these pairs, and a
parti ular solution sele ted. The values hosen were T1 = 0:336 and T2 = 0:144
whi h gave a drop-ba k of 0:1 s, an average phase rate of 57/Hz and f = 1:7
Hz. From Figure 27.12, it an be seen that this gives Level 1 performan e.
Figure 29.10 shows the time response of a pit h rate demand.

476
 1.6

1.4

1.2

0.8

0.6

0.4

0.2

0.2
0 0.5 1 1.5
time (s)

Figure 29.10: Step on q demand

29.4 Results from the Evaluation Pro edure

This hapter ontains results from the evaluation pro edure. These results are
from the rst appli ation of the pro edure, and indi ate that very little modi-
ation is required to the ontroller in order to meet all of the requirements. The
authors believe that this is in part due to one of the key strengths of the design
method in that loop-shaping gives good visibility as to what performan e an
be a hieved from the system. Hen e, the design is likely to exploit the sys-
tem's inherent performan e limits fully. At the same time, robustness must be
maintained so as to respe t the Ni hols ex lusion regions. This is equivalent to
a hieving a large enough value of .
It should also be borne in mind that the ontroller designed for ini3005 is
used at all operating points in the analysis, whereas in pra ti e it would be gain
s heduled with airspeed. Note that the ontrol law tested here is non-linear in
that it ontains models of a tuator rate and authority limits, and is in dis rete
time. Hen e it is very realisti and implementable.
All of the Ni hols plot tests were satised, with the ex eption of the one
shown in Figure 29.11 where the ex lusion region is just lipped. A small
in rease in the phase advan e in the dierential taileron weight is required.
The results indi ate a suitably robust design a ross the envelope despite not
gain s heduling the ontroller.
Figure 29.12 shows the Gibson Ni hols plot riterion for small perturbations
of the sti k. It an be seen that the response is Level 1. The Ni hols plot for
large sti k demands has not been produ ed yet. The drop-ba k and phase rate
riteria are given in Table 29.1. All meet the Level 1 requirements. Table 29.2
gives data on the response to turbulen e.
Figures 29.13, 29.14, 29.15 and 29.16 show responses resulting from om-
mands in pit h, roll, side-slip and speed. Dierent ight ases have been used
to demonstrate that performan e is good a ross the ight envelope. In Figure
29.13 a 5 /s pit h rate ommand is made. Note the pit h rate overshoot used
to give the required drop-ba k. At 4 s the in iden e limit is hit, and the alpha
limiter for es the pit h rate down, so as to maintain alpha on the limit. Note
also that the oupling into side-slip at high alpha is well ontrolled. In Figure

477
 Flight ondition Drop-ba k f _
Ma h 0.20,1000 ft 0:17 s 1.32 56:0/Hz
Ma h 0.24,20000 ft 0:23 s 1.40 56:0/Hz
Ma h 0.30,5000 ft 0:10 s 1.58 56:6/Hz
Ma h 0.40,10000 ft 0:07 s 1.40 57:0/Hz
Ma h 0.50,15000 ft 0:12 s 1.35 57:1/Hz

Table 29.1: Drop-ba k and phase rate riterion

RMS roll rate due to turbulen e 1.7272

/s
RMS pit h rate due to turbulen e

0.2470 /s
RMS yaw rate due to turbulen e

2.1431 /s
RMS normal a eleration due to turbulen e 0.1317 g

Table 29.2: Response to turbulen e

29.14 a maximum roll-rate ommand is made. Performan e is reasonable in

that this is a severe manoeuvre whi h results in the air raft banking over well
past 90. Coupling to side-slip is small (about 1 ), and oupling to q small in
omparison to the demand on p.
In Figure 29.15 a 10 step on side-slip is ommanded. The response time
is well within the requirements. De oupling from other ontrolled outputs is
also good. In Figure 29.16 a rapid speed demand hange is made. Note how
full use of the engine is made, and that the ontrol law respe ts the engine
rate limits. However there is some speed overshoot whi h should be redu ed,
possibly by redu ing the ross-over frequen y of the speed loop, or by limiting
the dynami s of the speed demand.
Overall, reasonable performan e was a hieved from the design. As with any
design approa h, a degree of renement will be possible with some iteration.
However our design ame lose to meeting all requirements on the rst attempt,
and this was with very limited time available to the authors.

29.5 Con lusions

A stability augmentation system has been developed for the HIRM using H1
loop-shaping. In the authors' view, the design problem was relatively simple,
parti ularly given the narrow ight envelope of the HIRM. However, the ap-
proa h whi h has been taken will extend in an easy manner to a mu h wider
ight envelope. Key to this is the observer stru ture used to implement the
H1 ontroller. The ontrol law stru ture has mu h in ommon with a lassi al
ontrol law, the main dieren e being the addition of the lateral and longitu-
dinal H1 ontrollers to the feedba k path. The observer stru ture is also used
to handle a tuator authority and rate limiting in a systemati fashion. Ap-

478
 Nominal flight, Nichols plot, Mach 0.24, 20000 ft
20

15 dts

dtd
10
o dcs
OpenLoop Gain (dB)

x dcd
5
+ dr

0 * thr

10

15

20
350 300 250 200 150 100 50 0
OpenLoop Phase (deg)

Figure 29.11: Ni hols plot of the nominal plant at Ma h 0.24, 20000 ft

Gibson criterion, pitch: linear analysis

20

15

10
Gain (dB)

5
L1

5
L1

10

180 160 140 120 100 80 60 40

Phase (deg)

Figure 29.12: Gibson Level 1 boundaries

479
 pw q az
0.5 10 8

9
0 5

(m/s^2)
(deg/s)

(deg/s)
10
0.5 0
11

1 5 12
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

Va alpha beta
70 35 0.4

68 30
0.2
(deg)

(deg)
(m/s)

66 25
0
64 20

62 15 0.2
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

q cmd dts dtd

6 10 4

0
4 2
(deg/s)

(deg)

(deg)

10
2 0
20

0 30 2
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

dcs dcd dr
20 1 15

10 0.5 10
(deg)

(deg)

(deg)

0 0 5

10 0.5 0

20 1 5
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

thr1 thr2
100 100
(%)

(%)

50 50

0 0
0 5 10 0 5 10
time (s) time (s)

Assessment manoeuvre: pitch rate demand, Mach 0.2, 1000 ft

Figure 29.13: 5/s pit h rate demand

480
 pw q az
60 2 5

40 0
0

(m/s^2)
(deg/s)

(deg/s)
20 2
5
0 4

20 6 10
0 2 4 6 0 2 4 6 0 2 4 6
time (s) time (s) time (s)

Va alpha beta
106 15 2

1
104 10
(deg)

(deg)
(m/s)

0
102 5
1

100 0 2
0 2 4 6 0 2 4 6 0 2 4 6
time (s) time (s) time (s)

p cmd dts dtd

80 0 40

60 20
5
(deg/s)

(deg)

(deg)

40 0
10
20 20

0 15 40
0 2 4 6 0 2 4 6 0 2 4 6
time (s) time (s) time (s)

dcs dcd dr
10 1 20

0.5 10
0
(deg)

(deg)

(deg)

0 0
10
0.5 10

20 1 20
0 2 4 6 0 2 4 6 0 2 4 6
time (s) time (s) time (s)

thr1 thr2
6 6

4 4
(%)

(%)

2 2

0 0
0 2 4 6 0 2 4 6
time (s) time (s)

Assessment manoeuvre: roll rate demand, Mach 0.3, 5000 ft

Figure 29.14: 70/s roll rate demand

481
 pw q az
1 0.05 9.4

0.5
0

(m/s^2)
(deg/s)

(deg/s)
0 9.6
0.05
0.5

1 0.1 9.8
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

Va alpha beta
161.4 7.1 15

10
161.2
(deg)

(deg)
(m/s)

7 5
161
0

160.8 6.9 5
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

betcmd dts dtd

10 5

6.6
0
(deg)

(deg)

(deg)

5
5
6.7

0 10
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

dcs dcd dr
0.1 1 15

0.5 10
0.05
(deg)

(deg)

(deg)

0 5
0
0.5 0

0.05 1 5
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

thr1 thr2
14 14

12 12
(%)

(%)

10 10

8 8

6 6
0 5 10 0 5 10
time (s) time (s)

Assessment manoeuvre: sideslip demand, Mach 0.5, 15000 ft

Figure 29.15: 10 side-slip demand

482
 x 10
10 pw q az
10 0.5 9

9.5
5 0

(m/s^2)
(deg/s)

(deg/s)
10
0 0.5
10.5

5 1 11
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
time (s) time (s) time (s)

Va alpha x 10
11 beta
180 12 3

160 10
2
(deg)

(deg)
(m/s)

140 8
1
120 6

100 4 0
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
time (s) time (s) time (s)

Va cmd dts x 10
10 dtd
160 4 1

5
140 0
(deg)

(deg)
(m/s)

6
120 1
7

100 8 2
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
time (s) time (s) time (s)

dcs dcd x 10
10 dr
1 1 10

0.5
0 5
(deg)

(deg)

(deg)

0
1 0
0.5

2 1 5
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
time (s) time (s) time (s)

thr1 thr2
100 100
(%)

(%)

50 50

0 0
0 10 20 30 40 0 10 20 30 40
time (s) time (s)

Assessment manoeuvre: air speed demand, Mach 0.3, 5000 ft

Figure 29.16: A 50 m/s speed hange

483
propriate handling of these non-linearities is essential for a hieving the desired
handling qualities, and ensuring sensible apportioning of ontrol power.
H1 loop-shaping is a fairly intuitive ontroller design method that an be
pi ked up in a relatively short amount of time, parti ularly from someone with
a ight ontrol ba kground. The design strategy employed bypassed the step
of trying to express in detail, the time response requirements in the frequen y
domain (a drawba k of frequen y domain based methods).
The order of the ontrol law is not seen as an issue as regards implementa-
tion provided that the omplexity an be justied in terms of desired robustness
and performan e. Experimentation with model redu tion showed that the lat-
eral H1 ontroller ould be redu ed to 10 states and the longitudinal to 8
states, both with minimal hange to the robustness and performan e. Bal-
an ed trun ation of a oprime fa torisation of the ontroller was used for this.
However, some are would be required when model redu ing a gain s heduled
design, as the physi al interpretation of the model redu ed states must be the
same for all designs to allow gain s heduling.
The type of un ertainty ( oprime fa tor un ertainty) adopted, although
quite general, did not prove to be too onservative. S heduling was avoided
partly be ause the open-loop HIRM was s aled with dynami pressure, a well
known te hnique within industry.
The paradigm of H1 loop-shaping is extremely powerful. There are other
extensions whi h have not been demonstrated here due to la k of time available
for the proje t. One ex iting new area is that of self-s heduled design meth-
ods whereby a parameter dependent ontroller is synthesised in one step for a
parameter dependent plant. For the HIRM, the obje tive would be to nd a
ontroller dependent on airspeed given a speed dependent model of the HIRM.
The synthesis of the ontroller relies on solving a set of linear matrix inequal-
ities for whi h there are numerous algorithms available. In the authors' view,
self-s heduled methods and linear parameter varying (LPV) plant des riptions
will provide very powerful and relevant tools for the aerospa e industry. In
parti ular, they provide a framework in whi h to address aerodynami non-
linearities and rate dependent ee ts. Further investigation of the best ontrol
stru ture is also a possibility. Designing a single ontroller for roll, pit h and
yaw would be worth arrying out to see what the ross-terms between lateral
and longitudinal feedba k loops are as a fun tion of ight ondition. This might
in lude looking at non-steady ight onditions su h as a non-zero roll rate.
For new types of ontrollers to be used by industry, there are two major
pre-requisites. Firstly benets need to be quantied so as to justify a hange in
approa h. Se ondly, on e the benets have been established, the ight lear-
an e aspe ts need to be addressed. The HIRM design hallenge was not set up
to address these issues dire tly. Its fo us was to demonstrate the methods to
industry, and to highlight what the advantages might be. As regards potential
benets of the method, there are two main ones worth mentioning. Firstly, for
omplex multi-input multi-output systems, the H1 loop-shaping method pro-
vides a way of synthesising a ontroller whi h ree ts the ross- oupling in the
plant. As su h, better performan e and robustness may be obtained sin e se-

484
le ting ontroller diagonal terms with non-multivariable methods an be a trial
and error task. Se ondly, the H1 loop-shaping approa h has great potential
to redu e design time. This is partly a result of its ability to handle om-
plex multivariable systems, and partly that given the nature of the robustness
optimisation, it is hard to design a bad ontroller.
Flight learan e of this type of ontrol law may be easier than for many
other non- lassi al design methods. In the main, this is be ause the loop-
shaping aspe t has dire t onne tions to lassi al design. The main dieren e
to a lassi al design is the addition of the multivariable H1 feedba k ontroller
element. However, the H1 feedba k element is implemented in an observer
form for whi h the stru ture is lear, and for whi h interpretation of the physi al
units of individual gains are lear. However, this is still (in the UK and probably
many other ountries) a new approa h to have state-spa e elements within
the ontrol law as opposed to SISO transfer fun tions. However, this will
probably not be an insurmountable problem, parti ularly given the numeri al
and e ien y advantages of using state-spa e implementations. The next step
therefore has to be to quantify benets in some meaningful manner.

485
30. Design of Stability Augmentation
System using -Synthesis

Karin Sthl Gunnarsson 1

Abstra t. It is des ribed how the -synthesis method was used to

design a ontroller for the HIRM ben hmark problem. Controllers
for the longitudinal and lateral axes were designed. For the longi-
tudinal axis a xed-gain ontroller was designed, but for the lateral
axis designs were arried out for a low and a high angle of atta k
ight ase. The main on lusions on erning the use of -synthesis
are: Handling qualities requirements as formulated in the HIRM
problem an easily be in orporated in the design. Robustness to
modelling errors an be handled in a straightforward way whereas
robustness due to variations in ight ase an be more di ult to
in lude. It may be a bit di ult to hoose weighting transfer fun -
tions that ree t the requirements in an appropriate way, and quite
some time is spent on tuning the weighting transfer fun tions.

30.1 Introdu tion

The following se tions des ribe the design and analysis of a ight ontrol system
for the HIRM, see 27. Sin e the requirements on the ontrol system emphasize
robustness and de oupling, the use of a multivariable robust design method
is motivated. The approa h hosen in this hapter is the -synthesis te h-
nique, see 8, [60 and [59, whi h gives the designer a dire t way to in orporate
robustness and de oupling aspe ts into the design pro ess.
In Se tion 30.2 the ontrol stru ture is presented. Next, in Se tion 30.3
the translation of the HIRM design riteria into an inter onne tion stru ture
is dis ussed. Then the design y le is des ribed in Se tion 30.4. Se tion 30.5
treats a - analysis of how the requirements are fullled and in Se tion 30.6,
the results from the automated evaluation are given. Finally, some on lusions
are given in Se tion 30.7.

30.2 Sele tion of Controller Ar hite ture

The approa h taken here is to split the ontroller into two parts: a longitudinal
and a lateral/dire tional ontroller.

1
Saab Military Air raft, S-581 88 Linkping, Sweden

486
 The longitudinal ontroller onsists of a linear part providing robustness
and handling qualities. The linear ontroller of the pit h axis has xed gains,
and is designed to operate over the omplete ight envelope. Longitudinal
sti k dee tion and throttle position are transformed to pit h rate and velo -
ity ommands respe tively. The ontroller uses velo ity V, pit h rate q and
normal load fa tor az for feedba k. The ontroller outputs symmetri taileron
dee tion T S and symmetri anard dee tions CS demands, as well as right
and left throttle position demands T H 1 and T H 2 . The intention has been
to use both symmetri taileron and symmetri anard for manoeuvering, with
only symmetri taileron being used for trimming. The engines are onsidered
to operate symmetri ally.
In addition to the linear ontroller there is a manoeuvering load limit (MLL)
fun tion blo k whose purpose is to limit the angle of atta k and load fa tor at
10o and +30o angle of atta k and 3g and +7g normal load fa tor. There are
also nonlinear ompensations of the symmetri anard dee tions and throttle
positions.
The lateral ontroller onsists of a linear part and a blo k al ulating the
gravity ompensation for yaw rate. The linear ontroller onsists of one on-
troller for low angle of atta k onditions and one ontroller for high angle of
atta k onditions. In the intermediate angle of atta k region there is a blending
of the outputs of the two ontrollers. The lateral sti k dee tion and rudder
pedal signals are transformed to velo ity ve tor roll rate and sideslip ommands
respe tively. The ontroller uses roll rate p, yaw rate r, roll angle  , sideslip
 , pit h attitude  and speed V . Roll angle, pit h attitude and speed are used
to reate the signal
g
r0 = r sin() os( )
V
whi h is used to oordinate the turn. The lateral ontroller outputs dierential
taileron dee tion T D and rudder dee tion R demands. Dierential anard
dee tions are not used sin e it is usually not very ee tive to use for roll
manoeuvres.
The ontrollers resulting from -synthesis are generally of high order even
after the use of model redu tion s hemes. In this ase the order of the longitu-
dinal ontroller is 13 and the lateral ontroller is of order 13. The ontrollers
are stable. The longitudinal ontroller in ludes two integrators.

30.3 Translation of Design Criteria into Method

Dependent Obje tives
This se tion des ribes how the dierent robustness and performan e design
riteria of the HIRM problem dened in 27.3 are transformed into an inter on-
ne tion stru ture that will be used in the -synthesis design pro edure. The
inter onne tion stru ture is shown in Figure 30.3 and onsists of the air raft
model augmented with un ertainty models, ideal response models and weight-
ing fun tions. If the un ertainty models
r ,
and the ontroller K are "pulled

487
 MLL-
function
+
+
qc
dTS
A B M,dTS
Vc +
V
dCS

q - + C D
az
0.5 dTH1
- Linear controller dTH2

Vtrim
aztrim

Figure 30.1: Stru ture of the pit h ontroller

pc A B dTD
c

p + C D dR
r

- Linear controller
g/Vsin()cos()

Figure 30.2: Stru ture of the lateral ontroller

out of the inter onne tion stru ture, one gets the general problem des ription
shown in 8.3. The three input-output pairs are the ontrols and the mea-
surements (u; y), the disturban es and the errors (d; e) and the perturbation
signals (w; z ). Here d = (d md ; dnoise ; dgust ), e = (eperf ; ea t ), w = (wr ; w )
and z = (zr ; z ). Below the dierent elements of the inter onne tion stru ture
are des ribed.

The elements of the inter onne tion stru ture that onstitute the air raft,
a tuator and sensor models are the blo ks HIRM, A tuator and Sensor. The
rigid body air raft model is ontained in the blo k HIRM and onsists of a
linear state-spa e model of the longitudinal or lateral dynami s at the design
ight ase. In the longitudinal ase, the engine dynami s are in luded as well.
The a tuator models are found in the blo k A tuator and onsist of the linear
a tuator models as given in Se tion 27.2.4. The blo k Sensor onsists of an
approximation of the sensor dynami s. The sensor dynami s are approximated
with a rst order Pad model orresponding to a time delay of 60 ms.

488
 e act e perf

Wact Wperf

Ideal
model

zc wc
c
wr zr
r
Wdel
Wr Wl

d cmd
Wcmd K Actuator Hirm
u

d noise
Wnoise Sensor

d gust
Wgust

Figure 30.3: Inter onne tion stru ture

The un ertainty model, a ording to the problem denition, overs both

the variation of the HIRM model due to dierent operating points in the ight
envelope and modelling errors. The variations due to operating points hanges
were modelled as an unstru tured omplex multipli ative un ertainty,
, at
the input of the air raft model, while the modelling errors were modelled by a
stru tured real un ertainty,
r , in the air rafts state-spa e model. The size of
the un ertainties are s aled by Wdel , Wr and Wl .
The elements of the inter onne tion stru ture that dene the performan e
model are ontained in the blo ks Ideal model, W md, Wnoise , Wgust , Wperf
and Wa t .
The blo k Ideal model ontains the ideal response models. These are hosen
to ree t the desired handling qualities. In the longitudinal ase, the dierent
pit h axis requirements are in luded in the design by the reation of an ideal
pit h rate response model. In this ase it was possible to dene a model su h
that both the pit h attitude frequen y response requirements as well as the
pit h rate time response requirements were fullled. The transfer fun tion
from sti k dee tion to pit h rate that was found to full all requirements is:

q 0:83s + 1
=9
q (0:05s + 1)(s + 4:8s + 9)
The motivation for the hoi e of this transfer fun tion is the following: From
nz = 4:8 1=rad
a simulation at Ma h 0.3 and altitude 5000 ft the value of
 is
found. From [2 a short period frequen y value wsp of about 3rad=s is obtained.
A damping ratio sp  0:8 was onsidered to be reasonable. In order to full
the frequen y response requirement a term giving some extra phase lag was
introdu ed.
In this design only the requirements for the speed ontroller for small throt-
tle travel is taken into a ount. The requirement leads to the following hoi e

489
of ideal speed response model:

v 1
=
v 1 + s
The ideal response models for pit h rate and speed respe tively are diagonally
augmented. This stru ture of the ideal response model gives de oupling be-
tween pit h rate and speed.
The lateral requirements are also taken into a ount by dening the roll
rate and sideslip ideal response models. The roll rate requirement leads to the
following hoi e of ideal roll rate response model:

p 1
=
p 1 + 0:4s
For the sideslip demand system the requirement is satised by the following
hoi e of ideal sideslip response model:

 1
=
 1 + s
The ideal response models for roll rate and sideslip respe tively are also diag-
onally augmented.
The signals that onstitute the overall design obje tives are then reated.
These error- or performan e variables onsist of tra king errors, a tuator de-
e tions and rates. The tra king errors are the dieren e between the ideal
response and the a tual signal and are weighted by Wperf . The inverse of this
weight indi ates how large the allowed steady-state tra king error due to om-
mand inputs is. The a tuator dee tions and rates are weighted with Wa t .
The weightings an be thought of as being about the inverse of the maximal
allowable value. The weight of the ontrol surfa e rates an be used as a tool
to ontrol the bandwidth of the system.
d md ,
The disturban e signals ae ting the errors onsist of pilot ommands,
sensor noises, dnoise dgust . The magnitude and frequen y on-
and wind gusts
tent of these signals are shaped with W md , Wnoise and Wgust .

30.4 Controller Derivation - Design Cy le

30.4.1 Longitudinal ontroller design
The approa h taken for the former is to perform the linear design for the ight
ase Ma h 0.3 and altitude 5000 ft and make the design robust enough to fulll
the requirements for all ight ases.
The rst step in the design y le is to dene numeri al values of the transfer
fun tions in the inter onne tion stru ture dened in Figure 30.3 in Se tion 30.3.
To reate the air raft model ontained in the blo k HIRM, ve states, ( , u
w, q , engine 1, engine 2) ve inputs, ( T S , CS , T H , WXE WZE ,) and three
outputs, ( V , q, az ) were sele ted. Note that T H = T H 1 + T H 2 , i.e. the thrust

490
 ommand used in the longitudinal model, is the sum of the individual engine
ommands. WXE and WZE are the axial and verti al wind gust omponents.
Next, inputs and outputs are added in order to model un ertainties in the
oe ients of the linear state spa e model. In order to study ee ts of vari-
ations in the aerodynami stability derivatives Cm ; Cmq ; CmT S and CmCS ,
four extra inputs and outputs, zr and wr , were added to the nominal system
des ription.
The weighting fun tions of the un ertainty model Wdel , Wr and Wl are
then dened. The un ertainties due to operation point variations around the
nominal ight ase at Ma h 0.3 and altitude 5000 ft, were found to be about
60 % for all frequen ies. This gives Wdel (s) = diag[0:6; 0:6; 0:6. The sizes of
the stru tured un ertainties are 10 % of their nominal value for all un ertain-
ties ex ept for the rst one. For this oe ient the absolute error of 0.001
is re al ulated into a value in per ent using the nominal value. This gives
Wl = [abs(0:001=A(5; 3)); 0:1; 0:1; 0:1, where A(5,3) is the nominal value
diag
of the oe ient in the state-spa e model. Wr is taken as an identity matrix.
In the longitudinal design model, the disturban es signals onsist of wind
gusts, sensor noises added to the measured signals and the pilot ommands.
In order to ree t the size and frequen y ontent of these signals, weighting
fun tions are hosen.
The wind gust weighting fun tion is Wgust = [WWXE (s); WWZE (s)
diag
where
+1 s
WWXE (s) = WWZE (s) = 2
s+1
ree ting a frequen y ontent of 1 rad/s and a magnitude of 1 m/s. The weight-
ing fun tion of the sensor noise is W =
noise diag v q [W (s); W (s); W
az . The (s)
dierent sensor noise omponents are supposed to be given all by the same
weighting fun tion:

s +1
Wv = Wq (s) = Waz (s) = 0:003 10s
200 + 1
The load fa tor weighting fun tion was then in reased by a fa tor 1000. The
pilot ommands are pit h rate, q V . The frequen y
and speed, ontent of
these signals are modelled by W md = diag[Wq (s); WV (s), where

s +1 s +1
Wq (s) = 0:2 20s WV (s) = 10 2s
5 +1 0 :5 + 1
The bandwidth of the pilot ommands in pit h rate and speed, are assumed
to be about 5 rad/s and 0.5 rad/s respe tively. The maximal amplitude of the
ommands are assumed to be 0.2 rad/s and 10 m/s respe tively.
The ideal models of the pit h rate and the speed ommand responses were
dened in Se tion 30.3. The pit h rate and speed tra king errors are weighted
by Wperf = diag[Wqerr (s); Wverr (s) where
s +1 s
Wqerr (s) = 500 300 30 + 1
s + 1 Wverr (s) = 50 s + 1
0 :3 0:03
491
This means, the pit h rate ommands should be followed with an a ura y
of 0.002 rad/s for low frequen ies, while the requirement is relaxed for higher
frequen ies. Emphasising on the tra king requirement results in integral a tion
to be in luded in the ontroller. The requirement on the speed response is not
as high. The ommand should be followed with an a ura y of 0.02 m/s at low
frequen ies.
The dee tions and rates of the a tuator signals should be minimized. This
is done by weighting the taileron dee tion and rate and anard dee tion and
rate and throttle dee tion with Wa t = diag[WT S ; W_ T S ; WCS ; W_ CS ; WT H
where

s +1
WT S (s) = 0:1 0s:5 W_ T S (s) = 1:5
50 + 1
( s + 1)( 2s + 1)
WCS (s) = 10 s0:5 s + 1) W_ CS (s) = 1:5
( 0:005 + 1)( 200
( s + 1)(s + 1)
WT H (s) = s0:1 s + 1)
( 0:01 + 1)( 100
The rate weighting fun tions are onstant, while the dee tion weighting fun -
tions are frequen y dependant. In order to avoid the use of the anard for
trim, the anard dee tion weighting fun tion as well as the thrust weighting
fun tion are large for low frequen ies, small for intermediate frequen ies and
large for high frequen ies.
Now, the inter onne tion stru ture of the longitudinal design problem is
dened. If the
's and the K are pulled out, the open-loop stru ture P,
with three pairs of inputs and outputs is a hieved. Here, the inputs and out-
puts that orrespond to the un ertainty hannel are given by z = [zr ; z ,
w = [wr ; w . The inputs and outputs to the performan e hannel are d =
[WXE ; WZE ; V ; q ; az ; q ; V , e = [qerr ; verr ; T S ; _T S ; CS ; _CS ; T H , The mea-
surement/ ontrol hannel are given by y = [q ; V ; V; q; az , u = [T S ; CS ; T H .
In this ase, it was not found ne essary to in lude the stru tured un er-
tainties in the design, sin e the design be omes robust enough anyway. Sin e
the inter onne tion stru ture in ludes the un ertainty des ription for analysis
purposes, the system is redu ed by taking away the four inputs and outputs
that orrespond to the un ertainty des ription, that is the input output pair
(zr ; wr ).
The next step in the design pro ess is to nd a ontroller, K , that fullls the
design obje tives by using the D K -iteration pro edure, see 8.4. In the rst
D K -iteration, an H1 ontroller is a hieved with all D-s alings equal to unity.
The optimal value of a hieved in the rst iteration is 18.15. The -value of the
losed-loop system is then al ulated together with the frequen y dependent D -
s alings. Low order transfer fun tions are tted to these frequen y responses.
The inter onne tion stru ture is then augmented with the transfer fun tions
D and D 1 . In this ase we have four un ertainty blo ks: three orresponding
to the unstru tured un ertainties and one orresponding to the performan e
blo k. This implies that three D-s aling transfer fun tions should be found

492
sin e the last blo k in D is kept to I . Se ond order transfer fun tions were
hosen for all D-s alings.
Now, a new H1 ontroller an be designed. After three iterations the redu -
tion of leveled o. The result of ea h iteration is summarized in Table 30.1.
The nal a hieved was 2.91, whi h implies that the performan e requirements
are not a hieved.

Table 30.1: Summary of longitudinal ontroller design

Iteration Order of system value a hieved

1 31 18.15
2 43 2.91
3 43 2.88
4 43 2.91

In Figure 30.4 the maximum singular value of the frequen y response of the
losed-loop system,  (M ), where M = M (P; K ) dened in 8.3, and the maxi-
mum value of the stru tured singular value,(M ) after the fourth iteration are
shown. Note that if (M ) had been less than  (M ) for all frequen ies would
have indi ated that additional D -s alings would have been useful for further
optimisation of the design. Sin e the -values in Figure 30.4 are larger than
1, this indi ates that the design obje tives are not a hieved. However, from
time simulations it was found that the properties of the ontroller is quite sat-
isfa tory. This indi ates that some of the weighting fun tions do not ree t the
a tual requirements to full extent and should be hanged. It ould also indi ate
that the design method is onservative or that it an not take all the a tual
requirements into a ount without making the design too onservative. How-
ever, due to the limited time available for this study, no hanges of weighting
fun tions were arried out.

3 14
sigma mu with total controller
mu mu with truncated controller
12
2.5

10

2
8
mu

mu

6
1.5

1
2

0.5 3 2 1 0 1 2 3
0 3 2 1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10 10 10
Frequency [rad/s] Frequency [rad/s]

Figure 30.4: -analysis of longitu- Figure 30.5: -analysis of total and

dinal design trun ated ontroller

The resulting ontroller after this design has order 43. It was redu ed using

493
standard model redu tion te hniques. The order of the redu ed ontroller was
13. In Figure 30.5 the -values showing the robust performan e level for the
ontrollers of order 43 and 13 respe tively are shown. As an be seen in the
gure, the -value in reases at low frequen ies for the 13-th order ontroller.
This ontroller was hosen anyway, sin e time simulations did not show signif-
i ant performan e deterioration when ompared to the 43-th order ontroller
as is shown in Figure 30.6.

14 0.5

12
0

10
0.5

8
q [deg/s]

V [m/s]
1
6

1.5
4

2
2

0 2.5
0 2 4 6 0 2 4 6
Time [s] Time [s]

0 5

1
4
2

3 3

4
dCS [deg]
dTS [deg]

2
5
1
6

7 0
8
1
9

10 2
0 2 4 6 0 2 4 6
Time [s] Time [s]

Figure 30.6: Time simulations with total (-) and trun ated (- -) ontroller.

30.4.2 Lateral ontroller design

The linear design was again performed for the ight ase Ma h 0.3 and altitude
5000 ft. The lateral air raft model in HIRM was reated by pi king out the
states, inputs and outputs appli able to the lateral hannel: three states ( , v p,
r), three inputs (T D , R , WY E ) and four outputs (p, r, ay,  ).
In order to des ribe un ertainties in the oe ients of the linear state spa e
model, extra inputs and outputs were added.
The pilot ommands are roll rate, p and sideslip,  . The frequen y ontent
and magnitude of these signals are shaped by W md = diag[Wp (s); W (s)
where
s +1 +1 s
Wp (s) = 1:25 20s W (s) = 0:2 2s
5 +1 0 :5 + 1
The roll rate ommand is given in velo ity axis and hen e the roll rate and

494
sideslip tra king errors are dened as:

perr = p os(o ) + rsin(o ) pideal err =  ideal

The error signals are weighted by Wperf = diag[Wperr (s); Werr (s) where
s +1 s +1
Wperr (s) = 50 300
s +1 Werr (s) = 200 30s
0:3 0 :3 + 1
In order to a hieve satisfa tory turn- oordination a modi ation of the yaw
rate r is performed
g
r0 = r sin() os( )
V
But the approa h taken here is to make the design as if r is the feedba k signal.
This probably results in an unstable spiral mode whi h an be a epted if the
time to double is long enough.
The dee tion and rates of the a tuator signals should be minimised. This is
done by weighting the taileron and rudder dee tions with Wa t = diag[WT D (s);
WR (s) where
s +1 s +1
WT D (s) = 0:15 0:s35 WR (s) = 0:05 0:s35
35 + 1 35 + 1
In order to make the design robust for all ight ases, a robustness weighting
giving about 6 dB amplitude margin and 35 degrees phase margin has been
used to over unstru tured un ertainties. The weighting fun tion ree ting this
requirement is given by:
!
p10:60:62
Wdel (s) = p10:06:62
The size of the stru tured un ertainties is 10 % of their nominal value for all
un ertainties ex ept for the Cl and Cn derivatives. For these derivatives the
absolute error of 0.01 and 0.002 was re al ulated into values in per ent using
the nominal values.
However the stru tured un ertainties were not in luded in the design.
The ontroller design is arried out in a similar way to the longitudinal
ase. There are only two unstru tured un ertainties and therefore only two
D-s alings. The orders are hosen to be two for both transfer fun tions that
are tted to the D-s alings.
Four iterations were arried out. The results of the iterations are summa-
rized in Table 30.2.
In Figure 30.7 the -value showing the robust performan e of the resulting
ontroller is shown. Also in this ase the properties of the ontroller are quite
satisfa tory despite that the -analysis shows that the performan e obje tives
are not a hieved.
The ontroller resulting after this design has order 32. A redu ed order
ontroller of order 13 was al ulated. -analysis of the total and the trun ated
ontroller is shown in Figure 30.8.

495
 Table 30.2: Summary of lateral ontroller design

Iteration Order of system value a hieved

1 24 3.74
2 32 1.83
3 32 1.82
4 32 1.82

Another design was arried out for the ight ase at Ma h 0.24 and altitude
20000 ft. In order not to violate the ontrol surfa e dee tions limits, the
requirements on roll mode time onstant and sideslip due to roll rate ommands
were relaxed.

sigma
1.8 mu

2
1.6

1.4
mu

mu

1.5

1.2

1 mu with total controller, upper bound

1 mu with total controller, lower bound
mu with truncated controller, upper bound
0.8
mu with truncated controller, lower bound

2 1 0 1 2 3
0.6 2 1 0 1 2 3 10 10 10 10 10 10
10 10 10 10 10 10 Frequency [rad/s]
Frequency [rad/s]

Figure 30.7: -analysis of lateral Figure 30.8: -analysis of total and

design. trun ated lateral ontroller.

30.5 -Analysis of the Resulting Controller

Analysis of the properties of the resulting ontrollers was arried out for ve
ight onditions in the ight envelope of HIRM. The Ma h number and altitude
of the ight ases together with the symbol of the line type whi h represent
the ight ase throughout this se tion are given in Table 30.3 below.

30.5.1 Analysis of the longitudinal ontroller

For ea h of the ve ight ases -analysis has been arried out. In Figures 30.9
and 30.10 the results of the robust stability and nominal performan e tests are
shown.
If the system is robustly stable and fullls the nominal performan e require-
ments, the -value is less than one for all frequen ies. From the gures it is
found that at all ve ight ases the system is robustly stable, while nominal

496
 Table 30.3: Ma h number, altitude and line type for the ve ight ases

Ma h Altitude Line type

[ft
1 0.20 1000 -.
2 0.24 20000 - -
3 0.30 5000 -
4 0.40 10000 x
5 0.50 15000 o

0.35 80

70
0.3

60
0.25

50
0.2

40
Mu

Mu

0.15
30

0.1
20

0.05
10

0 3 2 1 0 1 2 3
0 3 2 1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10 10 10
Frequency [rad/s] Frequency [rad/s]

Figure 30.9: Robust stability tests Figure 30.10: Nominal performan e

of longitudinal system at dierent tests of longitudinal system at dif-
ight ases. ferent ight ases.

performan e is not a hieved. However for this design, the -value expressing
nominal performan e was not less than one for the design ight ase, whi h is
ight ase 3, sin e the weighting fun tions expressing the nominal performan e
were not well tuned. But sin e time simulations for the design ight ase look
satisfa tory, it an be expe ted that the same result is a hieved for ight ase
4 and 5 as well, sin e the -values for these ight ases do not dier a lot
from ight ase 3. It an however be expe ted that ight ase 1 and 2 will
exhibit poorer performan e, sin e the -values for these ight ases are larger.
That this is the ase is shown in Figure 30.11 and 30.12 that show the linear
responses to a pit h rate and a speed step ommand respe tively.

30.5.2 Analysis of the lateral ontroller

The properties of the lateral ontroller were also analysed. In Figure 30.13 and
30.14 the results of the -analysis arried out for the dierent ight ases are
shown. From Figure 30.13 it is found that all ight ases ex ept at Ma h 0.5
altitude 15000 ft fullls the robust stability requirement. The results of the
nominal performan e test are shown in Figure 30.14. This plot indi ates that
the low frequen y properties at the dierent ight ases diers from the design

497
 14 1.4

12 1.2

10 1
Pitch rate [deg/s]

Speed [m/s]
8 0.8

6 0.6

4 0.4

2 0.2

0 0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time [s] Time [s]

Figure 30.11: Responses to pit h Figure 30.12: Responses to speed

rate ommands for the ve ight ommand for the ve ight ases.
ases.

ight ase. That this is the ase, was veried by simulations.

80
1.2
70

1
60

0.8 50
Gain

40
Gain

0.6

30
0.4
20

0.2
10

0 3 0 3 2 1 0 1 2 3
2 1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10 10 10
Frequency [rad/s] Frequency [rad/s]

Figure 30.13: Robust stability test Figure 30.14: Nominal performan e

of lateral system for dierent ight test of lateral system for dierent
ases. ight ases.

30.6 Results of the Automated Evaluation Pro-

edure
In the automated evaluation pro edure dierent tests are arried out in order to
verify that the dierent requirements on erning robustness and performan e
are fullled. In this ase, the evaluation pro edure shows that the ontroller
resulting from this design exhibits quite satisfa tory properties.
The robustness properties are exemplied with Ni hols plots for a ight ase
at Ma h 0.4 and altitude 10000 ft. Ni hols plots for nominal system and with
parametri un ertainties are shown in Figures 30.15 and 30.16 respe tively.
From the gures it is found that the ex lusion region is not interse ted ex ept
for the symmetri anard loop at low frequen ies.

498
 Nominal flight, Nichols plot, Mach 0.4, 10000 ft Parametric uncertainty, Nichols plot, Mach 0.4, 10000 ft
20 20

15 dts 15 dts

dtd dtd
10 10
o dcs o dcs
OpenLoop Gain (dB)

OpenLoop Gain (dB)

x dcd x dcd
5 5
+ dr + dr

0 * thr 0 * thr

5 5

10 10

15 15

20 20
360 270 180 90 0 360 270 180 90 0
OpenLoop Phase (deg) OpenLoop Phase (deg)

Figure 30.15: Nominal system, Figure 30.16: Parametri un er-

Ni hols plot, Ma h 0.4 altitude tainty, Ni hols plot, Ma h 0.4 alti-
10000 ft tude 10000 ft

Figures 30.17 and 30.18 show the evaluation of the pit h and roll Gibson
riterion at Ma h 0.4 and altitude 10000 ft. From the gures it is found that
these requirements are fullled.

Gibson criterion, pitch, Mach 0.4, 10000 ft, 10% 30% Gibson criterion, roll, Mach 0.4, 10000 ft, 10% 30%
20 25

20
15
offset 15 offset
added added
10 10.8 dB 10 11.0 dB
10.4 dB 10.0 dB
5
Gain (dB)

Gain (db)

f (Hz) f (Hz)
5
L1 0.5 o 0 0.5 o
1.0 x Sluggish 1.0 x
2.0 + 5 PIO 2.0 +
0
3.0 * Oscillation 3.0 *
10

5
15
L1 Good
20
10
25
180 160 140 120 100 80 60 40 350 300 250 200 150 100 50 0
Phase (deg) Phase (deg)

Figure 30.17: Ni hols plots of fre- Figure 30.18: Ni hols plots of fre-
quen y response from sti k dee - quen y response from sti k dee -
tion to pit h attitude together with tion to roll attitude together with
requirement. requirement.

Some examples of non-linear time simulations are shown in Figures 30.19

to 30.21. Figures 30.19 and 30.20 show the responses to a pit h rate ommand
of 5 degrees/s at Ma h 0.3 altitude 5000 ft and Ma h 0.5 altitude 15000 ft
respe tively. Figure 30.21 shows the responses to a roll rate ommand of 60
degrees/s at Ma h 0.3 and altitude 5000 ft.

The evaluation pro edure also showed low ontrol a tivity due to turbulen e
and sensor noise.

499
 x 10
4 pw q az
6 10 5

4
5 10

(m/s^2)
(deg/s)

(deg/s)
2
0 15
0

2 5 20
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

Va alpha x 10
5 beta
101 25 2

100
20 1
(deg)

(deg)
(m/s)

99
15 0
98

97 10 1
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

q cmd dts x 10
4 dtd
6 0 5

4 5 0
(deg/s)

(deg)

(deg)

2 10 5

0 15 10
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

dcs dcd x 10
3 dr
5 1 2

0.5 1
0
(deg)

(deg)

(deg)

0 0
5
0.5 1

10 1 2
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

thr1 thr2
80 80

60 60
(%)

(%)

40 40

20 20

0 0
0 5 10 0 5 10
time (s) time (s)

Assessment manoeuvre: pitch rate demand, Mach 0.3, 5000 ft

Figure 30.19: Responses to pit h rate ommand at Ma h 0.3 altitude 5000 ft

500
 x 10
3 pw q az
1 10 5

0.5 10
5

(m/s^2)
(deg/s)

(deg/s)
0 15
0
0.5 20

1 5 25
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

Va alpha x 10
5 beta
162 15 4

160 2
(deg)

(deg)
(m/s)

10
158 0

156 5 2
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

q cmd dts x 10
3 dtd
6 4 2

6 1
4
(deg/s)

(deg)

(deg)

8 0
2
10 1

0 12 2
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

dcs dcd x 10
3 dr
5 1 1

0.5 0.5
0
(deg)

(deg)

(deg)

0 0
5
0.5 0.5

10 1 1
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

thr1 thr2
100 100
(%)

(%)

50 50

0 0
0 5 10 0 5 10
time (s) time (s)

Assessment manoeuvre: pitch rate demand, Mach 0.5, 15000 ft

Figure 30.20: Responses to pit h rate ommand at Ma h 0.5 altitude 15000 ft

501
 pw q az
100 10 9

5 9.5

(m/s^2)
(deg/s)

(deg/s)
50
0 10

0 5 10.5
0 2 4 6 0 2 4 6 0 2 4 6
time (s) time (s) time (s)

Va alpha beta
103 1

12 0.5
102
(deg)

(deg)
(m/s)

0
101
0.5
11.5
100 1
0 2 4 6 0 2 4 6 0 2 4 6
time (s) time (s) time (s)

p cmd dts dtd

80 4 20

60 10
6
(deg/s)

(deg)

(deg)

40 0
8
20 10

0 10 20
0 2 4 6 0 2 4 6 0 2 4 6
time (s) time (s) time (s)

dcs dcd dr
3 1 10

2 0.5 5
(deg)

(deg)

(deg)

1 0 0

0 0.5 5

1 1 10
0 2 4 6 0 2 4 6 0 2 4 6
time (s) time (s) time (s)

thr1 thr2
8 8

6 6
(%)

(%)

4 4

2 2

0 0
0 2 4 6 0 2 4 6
time (s) time (s)

Assessment manoeuvre: roll rate demand, Mach 0.3, 5000 ft

Figure 30.21: Responses to roll rate ommand at Ma h 0.3 altitude 5000 ft

502
30.7 Con lusions
The -synthesis method was applied to the HIRM ben hmark problem. It was
found that -synthesis an be used to design a ight ontrol system with good
handling and robustness properties.
It was found easier to design a longitudinal than a lateral ontroller. This
is however usually the ase. It was no problem to design a ontroller that
satised the requirements on roll rate response and sideslip response, but it
was more di ult to keep the sideslip small due to roll rate ommands. To
over ome this problem the s heduling of the lateral ontroller has to in rease,
or the inter onne tions stru ture must be hanged in order to better take this
requirement into a ount.
Another problem on erned the in lusion of integrator a tion into the on-
troller. Espe ially in the design of the longitudinal ontroller it is a benet
if the design of integrator a tion is in luded in the design pro ess instead of
adding integrators afterwards. However it was found that the way integrator
a tion was in luded in this design was not satisfa tory, sin e it seemed as if
more integrators than ne essary were in luded in the ontroller. A better way
to in lude integrators is to in lude integrators in the inter onne tion stru ture
and feed the output from the integrators to the ontroller. However, due to
time limitations this solution was not tried.
To make the ontrollers robust over the omplete ight envelope might
also ause problems. In the design of the pit h axis ontroller is was found
that it was su ient to des ribe the model variations due to varying ight
ase as a variation in gain at ea h input of the system. This approa h gave
a ontroller that works satisfa torily in almost the omplete ight envelope.
An alternative would be to des ribe the model variations over ight ases as
un ertainties in the dierent elements of the state-spa e model of the air raft.
However, onsidering the un ertainties as un orrelated will probably give a too
onservative result. Instead a linear parameter varying (LPV) model ould be
reated, see for instan e [260, [112. This results in a model with repeated
s alar un ertainties, but sin e the software tool that has been used, see [18,
does not support design with this kind of un ertainty model, this solution was
not tried.
-analysis was found to be a useful tool in the design pro edure. If the
weighting fun tions are properly tuned it an be a useful tool to test if the
requirements are fullled or not. In the latter ase, -analysis ould also be
used to investigate whi h requirement is driving the problem.
To summarize, the advantages of the -synthesis design method are:
 The same framework an be used both in the design and in parts of the
analysis of the ontroller.
 Requirements that an be expressed by linear transfer fun tions an easily
be handled by the method, as well as requirements on tra king error magnitude,
ontrol a tivity level and disturban e reje tion.
 Robustness onsiderations like modelling errors are straightforward to
in lude in the design.

503
  Within the framework it an be determined whi h requirement drives the
problem.
 Commer ially available software tools exist that an be used in the design.
 Parts of the design an be automated.
There are also disadvantages with the -synthesis design:
 Non-linear requirements annot be handled. This implies that require-
ments su h as angle of atta k and load fa tor limits and other non-linear re-
quirements annot be in luded in the design, but must be handled by separate
fun tions.
 Un ertainty models where the un ertainties annot be onsidered as in-
dependent annot be handled by the present version of the software tools.
 The resulting ontroller is often of high order.
 There is a need to hoose a large number of weighting transfer fun tions
whi h might be time onsuming for a non-experien ed designer.

The on lusion of this study is that the -synthesis method an be used to

design a ontrol augmentation system with good properties. It might however
be a problem to make the design robust over a large ight envelope due to prob-
lems in properly modelling un ertainties in the present version of the software
tool. In that ase gain-s heduling of  ontrollers must be onsidered. Alter-
natively an inner/outer-loop ontroller ould be used, where the inner-loop is
designed to equalize the plant over the ight envelope using for instan e dy-
nami inversion and where the outer-loop is designed by the -synthesis method
to give good handling and robustness properties, see [7, [200. Together these
methods form a powerful tool to ontrol un ertain nonlinear systems and should
be onsidered in future resear h.

504
31. Design of a Robust, S heduled
Controller using -Synthesis

Johan Markerink 1

Abstra t. In this hapter, H1 -, -synthesis is used to design

a ontroller for the HIRM. H1 -, -synthesis dire tly in orporates
both performan e and robustness obje tives in the design pro ess.
The shape of the losed-loop frequen y response is spe ied by the
designer with the use of frequen y dependent weighting fun tions
in the design set-up alled the inter onne tion stru ture. Using this
approa h two ontrollers for both the longitudinal and the lateral-
dire tional air raft motions were designed. Model redu tion was
used to redu e the order of these ontrollers to a total of 31 states.
The ontrollers were analyzed using H1 -, -analysis te hniques,
verifying omplian e with the design obje tives. Linear time sim-
ulation showed good robustness and performan e hara teristi s of
the resulting ontrolled air raft. To extend these hara teristi s
over a larger ight envelope, a ontroller output s heduling s heme
based upon dynami pressure was adopted. Several nonlinear ele-
ments were added to the ontroller to a ount for nonlinearities in
the dynami s and to implement limiting and departure prevention
requirements. Nonlinear test on the resulting losed-loop system
demonstrated that the ontroller provides the airplane with good
handling qualities over the designated ight envelope. Air raft re-
sponses show little or no deterioration when model un ertainties and
perturbations are used, proving the robustness of the ontroller.

31.1 Introdu tion

This hapter ree ts the design of a ontroller for the manual ontrol augmen-
tation problem of the High In iden e Resear h Model (HIRM) as dened in
hapter 27. This work served as the Delft University of Te hnology (DUT) en-
try for the HIRM design hallenge and was performed as a graduation proje t
by the author. The method used for this ben hmark is H1 -, -synthesis; one
of the robust ontrol methodologies re ently developed. Using this method,
ontrollers for both longitudinal and lateral-dire tional air raft motions are

1
Delft University of Te hnology, Fa ulty of Aerospa e Engineering, Kluyverweg 1, 2629
HS, Delft, The Netherlands. E-mail address: markerindutlsb3.lr.tudelft.nl or mark-
erinnlr.nl

505
designed. During this design, the ontroller ar hite ture and the method de-
pendent obje tives are hosen to omply with the HIRM design riteria as far
as possible.
The stru ture of this hapter is as follows: In se tion 31.2 the layout of the
hosen ontroller ar hite ture and a des ription of the internal ar hite ture is
given. Se tion 31.3 will be dealing with the translation of the HIRM design
riteria into the stru ture of a general H1 -, -synthesis design. This results
in the set-up of an inter onne tion stru ture in whi h the desired handling
qualities play a major role. Se tion 31.4 ontains a des ription of the design
y le and a motivation for some of the hoi es made during this design y le.
In se tion 31.5 the resulting linear ontrollers will be analyzed in both the
frequen y and the time domain. In se tion 31.6 the nonlinear results of the
method independent automated evaluation software will be reviewed. Finally,
in se tion 31.7, the on lusions and lessons learned are presented.

31.2 The Sele tion of the Controller Ar hite -

ture
This se tion deals with the layout of the ontroller ar hite ture used in the
HIRM design problem. Investigation of the linearized models provided with the
HIRM software at dierent points in the ight envelope, revealed that there is
little intera tion between the longitudinal and lateral-dire tional air raft mo-
tions. Therefore the - ontroller designs for the two sets of air raft motions are
performed separately, resulting in two - ontrollers and two separate ontroller
blo ks. For s heduling purposes, both blo ks have an additional input. This
input is used to s ale the ommanded ontrol surfa e dee tion with the inverse
of the dynami pressure. The s heduling parameter is onstru ted using:

6900
s heduling fa tor =
q
This means that the ight ondition with a dynami pressure of 6900 N/m
2
an be seen as nominal, requiring no s aling. For other ight onditions, the
hange in ontrol surfa e ee tiveness due to the hange in dynami pressure
is an elled by the s heduling fa tor. The sti k gains are hosen to ommand
12:5/s pit h rate and 90/s roll rate at the maximum sti k for es.

31.2.1 Longitudinal ontroller ar hite ture

For the longitudinal air raft motions, the proposed ontrol strategy alls for
a Pit h Rate Demand (PRD) system and a velo ity ve tor airspeed demand.
The longitudinal ontroller ar hite ture is depi ted in gure 31.1. In this ar-
hite ture the following blo ks are present:
 Dynami pressure s heduling: The ontrol surfa e ommands are multiplied
by the dynami pressure s heduling fa tor.

506
 dynamic pressure
scheduling

az_measured

"_measured
dts_demand

pitch rate cmd *

dcs_demand
Product

q_measured q-limiting
x' = Ax+Bu
y = Cx+Du
throttl1_demand
-controller +
air speed cmd
throttl2_demand

V_measured
+ throttle nonlinearity

V_trim

V-limiting

2_measured
f(u) 3
2_trim sin(2)-sin(2_tr)
Gain

Figure 31.1: Longitudinal ontroller ar hite ture

 q-limiting: The q-limiting blo k is a group of fun tions whi h is used to limit
the pit h rate ommand a ording to the manoeuvre limits. The normal
a eleration has to stay within the -3g to +7g region, while  may not
ex eed 10 and +30. The pilot's pit h rate ommand is redu ed if these
boundaries are approa hed or ex eeded.
 V-limiting blo k: The purpose of this blo k is to limit the high frequen y
ontent and the absolute value of the ommanded hange in airspeed. This
is performed by a se ond order lter and a feedba k of the throttle position
within this blo k.
 - ontroller: This is the a tual linear ontroller, represented by a state-spa e
system with 3 inputs and 3 outputs.
 Throttle nonlinearity: This blo k is used to resolve the nonlinear throttle
response explained in se tion 27.2.3. This nonlinearity results from the
dieren e between the dry and reheat thrust in rease as a fun tion of the
throttle ommand.
 Additional throttle demand: If the pit h attitude hanges, the gravitational
for e omponent from the mass of the airplane will indu e an additional for e
in the airplane X -axis. This for e will in turn indu e an unwanted hange
in airspeed. The additional throttle demand is designed to ompensate for a
hange in attitude by generating the following additional throttle demand:

additional throttle demand = KT H (sin() sin(trim ))

nonlinear testing revealed that in the dynami pit h manoeuvres onsidered,

KT H =3 was a good hoi e.

507
 roll rate cmd
dtd_demand

sideslip cmd 1
0.34s+1 dcd_demand
x' = Ax+Bu
Filter y = Cx+Du *
p_measured
-controller Product

$_measured dr_demand

dynamic pressure
scheduling

Figure 31.2: Lateral-dire tional ontroller ar hite ture

31.2.2 Lateral-dire tional ontroller ar hite ture

The lateral-dire tional ontrol strategy is that of a velo ity ve tor roll rate
demand, ommanded by lateral sti k and a sideslip angle demand, ommanded
by rudder pedal. In body axis, the velo ity ve tor roll rate is a fun tion of the
angle of atta k :
pW = p os() + r sin()
This poses a problem, be ause it is impossible to in orporate this nonlinear
relationship in a linear - ontroller. It would be possible to design a velo ity
ve tor roll rate system with -synthesis, but this would only be valid for one
angle of atta k. The design strategy for the - ontroller is therefore hanged
to ontrol body axis roll rate and sideslip. The sideslip tra king however,
is required to be very tight, so that a roll rate demand will only result in
small sideslip hanges. Nonlinear testing proved that this way, the ontroller
ee tively ontrols velo ity ve tor roll rate at any angle of atta k. To keep the
sideslip small, a body axis roll rate ommand automati ally also results in a yaw
rate, orresponding to pw sin( ).  The ar hite ture for the lateral-dire tional
ontroller is plotted in gure 31.2 and onsists of the following elements:
 Sideslip ommand lter: This lter is used to limit the high frequen y
ontent of sideslip ommands thereby alleviating any dynami al roll angle
oupling that might be present.
 - ontroller: This is the a tual linear ontroller, represented by a state-
spa e system with 4 inputs and 3 outputs. The ontroller designed with
H1 -, -synthesis is a 2 degrees-of-freedom ontroller. This means that the
ommands d and the measured outputs y enter the ontroller separately.
The - ontroller in orporates both stabilization as well as ommand shap-
ing, removing the need for an additional ommand path lter.
 Dynami pressure s heduling: The ontrol surfa e ommands are multiplied
by the dynami pressure s heduling fa tor.

31.2.3 Resulting ontroller omplexity and implementa-

tion issues
The omplexity of the resulting ontroller is mainly governed by the number
of states of the two - ontrollers. These are:

508
 frequen y [rad/s damping

9.8927e-006 1.0000e+000
5.0908e-003 1.0000e+000
8.6871e-002 1.0000e+000
1.1542e-001 1.0000e+000
4.5883e-001 1.0000e+000
3.9647e+000 1.0000e+000
1.7474e+001 4.9664e-001
1.9973e+001 1.0000e+000
5.2606e+001 2.4637e-001
9.4169e+001 5.6592e-001
2.2386e+002 7.0502e-001

Table 31.1: Frequen y and damping of the longitudinal - ontroller

pitch rate command measured pitch rate speed command

10 10 0
10 10 10
symm. taileron

0 0 2
10 10 10

10 10 4
10 5 0 5
10 5 0 5
10 5 0 5
10 10 10 10 10 10 10 10 10
10 10 0
10 10 10
symm. canard

0 0 2
10 10 10

10 10 4
10 5 0 5
10 5 0 5
10 5 0 5
10 10 10 10 10 10 10 10 10
5 5 5
10 10 10
throttle

0 0 0
10 10 10

5 5 5
10 5 0 5
10 5 0 5
10 5 0 5
10 10 10 10 10 10 10 10 10

Figure 31.3: Longitudinal - ontroller frequen y response

 longitudinal: 15 states
 lateral-dire tional: 16 states

This adds up to a total of 31 states for the - ontrollers. With other, more ad-
van ed model redu tion methods, this order ould possibly be redu ed further.
To ope with some of the nonlinear ee ts, three additional lters were used in
the ontroller adding 4 states. The total number of ontroller states is therefore
35. The longitudinal and lateral-dire tional - ontrollers are stable dynami
systems. Table 31.1 gives the frequen y and damping data of the longitudinal
- ontroller. To further enhan e the visibility, the frequen y responses from
inputs to outputs are plotted in gure 31.3.

509
31.3 The Translation of the HIRM Design Cri-
teria into -Obje tives
In this se tion the HIRM design riteria as des ribed in se tion 27.3 will be
translated into obje tives and riteria that an be used with an H1 -, -
synthesis design pro edure. To in lude the handling qualities riteria into the
-design, referen e models with ideal handling qualities are used. In se tion
31.3.1, these referen e models will be derived for both the longitudinal and
the lateral-dire tional air raft motions. Se tion 31.3.2 deals with the way in
whi h the robustness requirements are in orporated into the design. In se tion
31.3.3, the general inter onne tion stru ture in luding all the dynami s, lters
and weighting fun tions needed to perform the design, is onstru ted.

31.3.1 Handling qualities riteria

The longitudinal referen e model
The ontrol strategy for the longitudinal air raft motions of the HIRM alls
for a Pit h Rate Demand (PRD) system, ommanded by longitudinal sti k de-
e tion and a velo ity ve tor airspeed demand system, ommanded by throttle
lever. The longitudinal- ontroller to be designed will make the losed-loop
response, from pilot pit h rate ommands to resulting pit h rate, follow the
referen e model response as losely as possible. The PRD referen e model used
is a onstant speed equations of motion, se ond order transfer fun tion:

q 1 + q s
=
q 1 + 2!sp s + !s22
sp sp

This leaves three parameters to be hosen to onstru t the referen e model:

pit h mode time onstant:  q = 1 s.

short period natural frequen y: !sp = 1.67 rad/s.
short period damping:  sp = 0.7
These parameters were hosen using the following HIRM design riteria for
losed-loop response:
 Pit h attitude frequen y response riterion: The plot lies within the level 1
boundary for good handling qualities, as depi ted in gure 27.13.
 Dropba k riterion: The ratio of dropba k to steady state pit h rate, fol-
lowing a step input should lie between 0 and 0.25 se onds. The relationship
for dropba k is:
db 2sp
= q
qss !sp
For the given referen e model a dropba k of 0.16 se onds was a hieved.
For the speed demand system there are two riteria:
 Speed ontrol time onstant riterion: For small amplitude speed ommands
the time onstant should be in the range 0.75 to 1.5 se onds.

510
  Speed ontrol overshoot riterion: overshoot should be less than 3%.
To fulll these requirements, a rst order ideal speed ommand referen e model
with a time onstant of 1 se ond is hosen.

V 1 1
= =
V 1 + V s 1 + s
By diagonally augmenting the ideal Pit h Rate Demand system and speed
ommand system, one longitudinal referen e model is generated. This model
will be used in the set-up of the longitudinal inter onne tion stru ture used for
the H1 -, -synthesis design pro edure.
The lateral-dire tional referen e model
The ontrol strategy for the lateral-dire tional air raft motions of the HIRM
all for a velo ity ve tor Roll Rate Demand (RRD) system, ommanded by
lateral sti k dee tion and a sideslip demand system, ommanded by rudder
pedal dee tion.
The HIRM riteria on the Roll Rate Demand system are:
 Roll mode time onstant riterion: This time onstant should be p  0.4
se .
 Roll attitude frequen y response riterion: This riterion onsists of a Ni hols
plot from roll rate ommand (or lateral sti k dee tion) to roll angle. The
resulting graph should lie in the region labelled with good response in
gure 27.14.
To omply with the rst riterion a rst order model with the spe ied time
onstant is hosen. The rst order model was unable to generate su ient
high frequen y lag to stay within the boundaries. For this reason, a rst order
lowpass lter was added to provide the phase lag. The result is the following
roll mode referen e model:

p 1 1 1 1
= =
p 1 + p s 1 + lag s 1 + 0:4s 1 + 0:07s
For the sideslip demand system there is only one riterion:
 Sideslip step response riterion: The response of sideslip to a step input in
sideslip ommand should lie within the boundaries of gure 27.16.
The model used for the sideslip demand system is that of a simple rst order
transfer fun tion:
 1 1
= =
 1 +  s 1 + 0:5774s
The two lateral-dire tional referen e models are diagonally augmented to one
another, yielding one lateral-dire tional referen e model. This model is then
used in the set-up of the inter onne tion stru ture used for the H1 -, -synthesis
design pro edure.

31.3.2 Robustness onsiderations

The design ight envelope for the HIRM model is:

511
  Ma h: 0.15 to 0.5
 Altitude: 100 to 20000 ft.
In order to take the model variations over the ight envelope into a ount,
the linear system dynami s in various operating points are parameterized as a
fun tion of the dynami pressure. The method used in this design, is to model
the HIRM as a Linear Parametri ally Varying (LPV) system. Although the
linear representations of the nonlinear HIRM model are fun tions of several
variables, making them dependent on ight ondition and ight envelope, the
leading parameter is the dynami pressure q = 1/2V 2 . All aerodynami vari-
ables in the state-spa e model, vary linearly with q
. With the onstru tion
of the HIRM LPV-model, this linear dependen e is modelled via additional
inputs and outputs on the HIRM model. This means that in the -synthesis,
the variations of the model through the ight envelope are taken into a ount
through this varying parameter. For a more extensive review of the LPV model
and the onstru tion of this model in the HIRM design see the design report
[163.
In the HIRM manual, two additional kinds of modelling errors are des ribed:
errors on the aerodynami moment derivatives and errors on the total moment
oe ients. All these modelling errors are in orporated into the LPV HIRM
model as un ertainties on the individual matrix elements. These parametri
un ertainties are modelled using a real, stru tured, additive un ertainty repre-
sentation. As a result of this, the models are augmented with additional inputs
and outputs through whi h these un ertainties a t.
The gain and phase requirements di tate that the losed-loop system should
not be ome unstable when adding additional gain and phase osets at the input
of ea h a tuator demand. The denition of these requirements exa tly ts the
multipli ative un ertainty blo k used in this design, be ause the
m -blo k a ts
at the same point in the loop and adds its gain and phase osets. See gure
31.4.

31.3.3 Des ription of the inter onne tion stru ture

The inter onne tion stru ture is what needs to be dened prior to an a tual
-design. The layout of this stru ture, as well as the denition of the elements
within it, determines the su ess of a -design. The stru ture used in this
HIRM design is given in gure 31.4 and is valid for both the longitudinal and
the lateral-dire tional designs, but the exa t numeri al values of the elements
and fun tions will of ourse dier between the two.
The inter onne tion stru ture onsists of the following elements:
STICK This element represents the sti k lter. It is used to narrow down the
bandwidth of the pilot to more realisti values.
ID This element represents the referen e model with ideal handling quali-
ties. The synthesis pro edure will try to shape the losed-loop response
to t this referen e model.
K The H1 -, -synthesis ontroller. Before the design pro edure, the
ontroller is not yet present. Only the pla e in the inter onne tion

512
 e2
ID _
+
Wp

e1
wm zm
)m wa )qs za

Wdel Wacts
)a
cmd +
STICK K u +
ACTS HIRM
y

noise + +
Wnoise SENS

Figure 31.4: Inter onne tion stru ture

stru ture and the number and type of inputs and outputs is known.
Wdel The multipli ative robustness weight. This weighting fun tion s ales
the multipli ative perturbation-matrix
m . If Gnom is the nominal
plant, the set of plants hara terized by this un ertainty representation
is given by:

G
= Gnom (I +
m Wdel )
At any frequen y !, the magnitude of Wdel (!) an be interpreted as
the per entage of un ertainty in the model at that frequen y. For
example: a Wdel value of 0.5 represents a 50% modelling error at that
frequen y.

a The additive perturbation matrix
a is used to model the real, stru -
tured, additive un ertainties to the individual state-spa e matrix ele-
ments.

m The multipli ative perturbation matrix
m represents the input un-
ertainty in the model.

qs The perturbation matrix
qs is used to model the Linear Paramet-
ri ally Varying (LPV) nature of the HIRM-system. It introdu es the
linear dependen e of the A- and B -matrix elements on the dynami
pressure into the system.
ACTS This element represents the a tuator dynami s as des ribed in the
HIRM problem denition. The input is a tuator demand and the out-
puts are a tuator dee tion and rate. In the ase of multiple a tuators,
these are diagonally augmented to obtain one a tuator system.
Wa ts The a tuator weight. This weighting fun tion s ales the admissible
a tuator dee tions and rates to unity. In the simplest ase, Wa ts
would onsist of the inverse of the maximum a tuator outputs, result-
ing in an error (output of Wa ts ) of unity, if the maximum dee tion
was rea hed.
HIRM The linear, state-spa e model of the air raft.

513
 Wp The performan e weight. This weighting fun tion s ales the perfor-
man e error to unity. The input of Wp is the tra king error between
the response of the referen e model and the losed-loop system. This
weighting an be seen as the denition of the allowable deviations
from the referen e model response. Be ause the output errors are
s aled to unity, these deviations have to stay below the value of the
inverse of Wp at every frequen y. If at a ertain frequen y Wp =50,
this means that the losed-loop response may deviate 1/50 = 2% from
the referen e model.
SENS This element represents the dynami s of the sensors used. Also the
anti-aliasing, not h and averaging lters des ribed in the HIRM prob-
lem denition, are in orporated into SENS . As with the a tuator dy-
nami s, multiple systems are augmented diagonally into SENS .
Wnoise The noise weight. This weighting fun tion s ales the unity-intensity
noise entering the system. Be ause it represents high frequen y sensor
noise, high frequen y lters are used in Wnoise .
Breaking the loops at the ontroller K and at the perturbation
-blo ks and
olle ting the remaining systems into one system results in the generalized
plant P . The inputs and outputs of this generalized plant are depi ted in gure
31.5. The information of all systems and fun tions and the inter onne tions
between them, present in the generalized plant, is used by H1 -, -synthesis to
design a robust ontroller. The robustness and performan e properties of this
ontroller are related dire tly to the weighting fun tions used in the inter on-
ne tion stru ture.

{ }
wa za
w wm zm z

{ P }
cmd e1
d e2 e
noise

u y

Figure 31.5: Generalized plant

31.4 The Des ription of the Design Cy le

31.4.1 The general design y le
In the H1 -, -synthesis design for the HIRM design hallenge, several steps
need to be taken to ome to a ontroller that satises the design goals and
spe i ations. These steps will be referred to as the general design y le. The
larger part of this general design y le will be performed twi e, for the longi-
tudinal and lateral-dire tional designs separately.

514
 The general design y le onsists of the following steps:
1) Dene the design problem: the design problem denition states the overall
design problem to be solved. In this ase it onsists of the design of a
ontrol augmentation system that provides a high angle of atta k ghter
air raft with satisfa tory handling qualities. This design problem denition
is treated in hapter 27.
2) Translate the design problem into a general inter onne tion stru ture. This
stru ture denes the basi layout of plant, ontroller and ideal model inter-
onne tion.
3) Formulate riteria: a set of numeri ally dened riteria is used as spe i a-
tion of the overall design goals and to evaluate the resulting design. These
riteria are dened in se tion 27.4.
4) Translate handling qualities riteria into referen e models: for air raft on-
troller design problems in whi h handling qualities play a major role, like
the HIRM, a referen e model is an ideal method of translating the time do-
main riteria into riteria that an be used in an H1 -, -synthesis design.
The set-up of the referen e models with ideal handling qualities is treated
in se tion 31.3.1.
5) Translate robustness riteria into system perturbations: the ontroller has
to provide su ient stability and performan e over a range of ight ondi-
tions and has to be robust against several model un ertainties of the plant.
To a ommodate these riteria, the ontroller will be designed for a set
of plants. This set onsists of a nominal plant with several perturbations
a ting on it, to model the various un ertainties. The onstru tion of these
plant models and perturbations for the HIRM is treated in se tion 31.3.2.
6) Extend the general inter onne tion stru ture: several riteria an be im-
plemented dire tly into the design pro ess by adding elements, weighting
fun tions or inputs and outputs to the general inter onne tion stru ture.
The result of this pro ess an be found in gure 31.4.
7) Choose the weighting fun tions: after the inter onne tion stru ture is de-
ned, the only variable elements are the weighting fun tions. The hoi e
and onstru tion of these fun tions is the greater part of the design pro ess.
Usually numerous redesigns are ne essary to nd a ombination of weight-
ing fun tions that result in a ontroller whi h omplies with the riteria.
Weighting fun tions are in a sense, the knobs the designer an turn to
shape the resulting ontroller and losed-loop response.
8) Constru t the a tual inter onne tion stru ture and resulting generalized
plant: this step involves writing a omputer program that implements the
onne tion of all the models, lters and weighting fun tions. The output of
this program is a linear model of the generalized plant with the input/output
denition of gure 31.5.
9) Perform a -synthesis: using the generalized plant resulting from the previ-
ous step, the -synthesis pro edure designs the a tual ontroller. It onsists
of a D-K iteration s heme whi h involves an H1 -synthesis, a -analysis and
a tting of s aling fun tions.
10) Analyse the losed-loop system: using frequen y domain te hniques su h as

515
 Singular Value analysis and -analysis, the losed-loop system an be ana-
lyzed. Also, a variety of time domain te hniques an be used to investigate
omplian e of the resulting ontroller with the design riteria. If the design
riteria, frequen y or time domain, are not met, one has to go ba k to step
7), hange the weighting fun tions and redesign the ontroller. This pro ess
ontinues until all riteria are met or no more improvement is a hieved.
11) Redu e the order of the ontroller: the synthesized ontroller an have quite
a large number of states. To redu e the order of the ontroller a balan ed
realization of the ontroller is reated. This balan es the observability and
ontrollability grammians and orders the states a ordingly [50. Next a
trun ation an be used to remove (nearly) unobservable and un ontrollable
modes. In this design, a number of high frequen y modes that were in om-
patible with the HIRM's assumed 80 Hz Flight Control Computer (FCC)
iteration rate were removed also.
12) Analyse the losed-loop system: using the same te hniques as in step 10),
the losed-loop system with the redu ed ontroller an be analyzed. If re-
du tion deteriorates the ontroller performan e, one has to go ba k to step
11) and trun ate less states of the original ontroller. If the trun ated on-
troller exhibits no performan e dieren es with the original ontroller, one
an go ba k to 11) and try to trun ate the ontroller further. The ontroller
order redu tion ends when this balan e between ontroller performan e and
ontroller order has been found. A number of these analysis results for the
resulting ontrollers will be presented in se tion 31.5.
13) Constru t the ontroller ar hite ture: the - ontrollers are part of the om-
plete ontroller blo k. This blo k also onsists of input and output onne -
tions, signal additions and subtra tions and additional lters and fun tions.
In order to ope with nonlinearities in the model or in the riteria, some non-
linear and limiting blo ks may also be present. The set-up of this ontroller
ar hite ture is treated in se tion 31.2.
14) Test omplian e of the ontroller with the design riteria stated in 3): using
the full nonlinear model, the automated evaluation software he ks if the
ontroller meets these riteria. This nal analysis in ludes several nonlinear
time simulations, robustness and performan e he ks and other tests. Some
results of the evaluation an be found in se tion 31.6.

31.4.2 Pra ti al onsiderations of the design y le

In the numeri al exe ution of this design y le the matrix manipulation soft-
ware pa kage Matlab is used. In parti ular the add-on toolbox for analysis
and synthesis of robust ontrol systems -tools (mutools) is used extensively.
Besides the fun tions provided by mutools, additional programs written by the
author in Matlabs s ripting language, are used to implement ertain elements
of the design y le.
There is little intermediate analysis and redesign eort. Be ause the anal-
ysis part is in orporated into the design y le, it requires no extra eort. Both
frequen y and time domain analysis te hniques are used in this design y le.

516
 The frequen y domain analysis is strongly related to the used design meth-
ods: H1 -, -synthesis. These analysis te hniques an therefore be used to
see whether the design obje tives in terms of minimizing losed-loop gains are
a hieved. The linear time domain analysis has more onne tions with the a -
tual riteria, in parti ular, the riteria on erning handling qualities. If the
hoi e of the weighting fun tions and the set-up of the inter onne tion stru -
ture is done ideally, the frequen y and time domain analysis results oin ide.
Unfortunately this may prove to be very di ult to a hieve. The most reli-
able method of testing omplian e with the design obje tives and riteria uses
nonlinear time simulations. On e the ontroller ar hite ture is present, in-
termediate linear - ontrollers an be in orporated into this ar hite ture and
analyzed this way.
For a redesign after an air raft design hange, the part of the design y le
starting with step 7) has to be reperformed. The hoi e of new weighting
fun tions results in the loop from step 7) to step 10). After the riteria are
met, the order of the new ontroller an be redu ed with step 11) and 12).
The eort related to this redesign depends on the s ale of the hange. Certain
weighting fun tions may remain the same, while others may need onsiderable
hanging. If the air raft dynami s are hanged, the set-up of the HIRM model
in step 5) also needs to be reperformed resulting in a new LPV HIRM model.

0 3
10 10
pitch angle weight airspeed weight
2
10
1
10
1
pitch rate weight
10
gain

gain

2 0
10 10

1
10
3
taileron weight
10
2
10

4 3
10 4 2 0 2 4
10 4 2 0 2 4
10 10 10 10 10 10 10 10 10 10
frequency [rad/s] frequency [rad/s]

Figure 31.6: Robustness weight Figure 31.7: Performan e weight

31.4.3 Weighting fun tion sele tion

As stated earlier, the larger part of the design y le onsists of hoosing the
weighting fun tions, designing a ontroller, evaluating the resulting system and
redesigning the weighting fun tions; step 7) through 10). Be ause these weight-
ing fun tions are responsible for the hara teristi s of the resulting ontroller,
the su ess of the design depends upon their hoi e. Design dilemmas and
trade-os, su h as robustness versus performan e and performan e versus on-
trol a tivity are therefore ree ted in the weighting fun tions hoi e.
In the design y le used in the HIRM design, the design obje tives were to
minimize the robust performan e level as indi ated by the frequen y -plot.

517
A ording to the theory, if robust performan e is satised, Nominal Stability,
robust stability and nominal performan e are also satised. robust performan e
is a hieved if the -plot stays below unity over the entire frequen y range. It
is not un ommon to nd that in omplex designs this goal annot be a hieved.
With its many dierent and demanding riteria, design ight envelope and
several model variations, this is also the ase with the HIRM design. This does
not mean that the design will result in bad ontrollers. Remember that the
a eptability of the ontrol solution does not depend on  alone. Nevertheless,
the dis repan y in the theoreti al denition of a hieving design obje tives with
a -value below unity and the possibility of not a hieving this value is felt to
be a weak point of -theory. One riterion that had to be a hieved however,
was that of robust stability. The ontrollers had to guarantee stability for the
omplete set of model perturbations, as indi ated by a robust stability value
below unity at all frequen ies.
As an example of the use of weighting fun tions gures 31.6 and 31.7 give a
robustness and performan e weight of the longitudinal HIRM design.
Wdel For the taileron demand signal, the robustness weight is a rst order lter
with a 0.1% un ertainty at low frequen ies and 10% un ertainty at high
frequen ies. The motivation behind this hoi e is that the un ertainty of
the HIRM model is expe ted to in rease with frequen y.
Wp The longitudinal performan e weight onsists of a system with three di-
agonal entries: a weighting fun tion for the pit h rate error, for the pit h
angle error and for the speed error. These three weights are performan e
weights be ause they s ale the error between the losed-loop response and
the ideal referen e model response. The plot of the weighting fun tions
in gure 31.7 an be interpreted as the inverse of the allowable error. The
pit h rate error and pit h angle error weighting fun tions are rst order
low-pass lters. This puts the emphasis on steady state tra king and lim-
its the bandwidth in whi h the losed-loop system has to follow the ideal
models. The q -weight emphasizes that at low frequen ies errors may not
ex eed 10%. To a hieve some form of pit h attitude hold and pit h atti-
tude tra king in the nal losed-loop system, the pit h angle error is also
used as a performan e output. The pit h angle weight limits the pit h
angle error to 0.4%. The airspeed weighting fun tion has a large value
of 200 at high frequen y, limiting errors to 0.5%. This is to a hieve good
tra king of the transient part of the rst order ideal model. The weight is
in reased to 1000 at low frequen ies to ensure steady state airspeed hold
with minimal errors.

31.5 Analysis of the Resulting Controller

31.5.1 The frequen y domain analysis tests
The designs were analyzed using H1 -, -theory frequen y domain analysis
tools. First, the inter onne tion stru ture of gure 31.4 is onstru ted, resulting
in the generalized plant of gure 31.5. Closing the lower loop with the ontroller

518
8 7

7 Robust Performance: 6
Robust Performance:

6
5

5 Nominal Performance
4

4
3
3

Nominal Performance 2
2

Robust Stability 1
1 Robust Stability

0 -3 0 -3 -2 -1 0 1 2 3
10 10
-2 -1
10 10
0
10
1 2
10
3
10 10 10 10 10 10 10 10
frequency [rad/s] frequency [rad/s]

Figure 31.8: Frequen y analysis Figure 31.9: Frequen y analysis

longitudinal - ontroller lateral-dire tional - ontroller

K yields the losed-loop system. All analysis te hniques used in this se tion
will result in frequen y plots in whi h a value below unity indi ates a hievement
of the pertinent design goals. The meaning of the frequen y analysis plots is
as follows (see hapter 8 for the exa t denitions):
 Robust Stability: robust stability indi ates whether the system remains
stable for a given set of perturbations to the nominal model.
 Nominal Performan e: indi ates omplian e of the nominal, unperturbed
system with the design requirements for losed-loop performan e. These
requirements are implemented by the weighting fun tions in the inter on-
ne tion stru ture of gure 31.4.
 Robust Performan e: test if the losed-loop system meets performan e re-
quirements and remains stable for the given set of perturbations to the
nominal model. This test is performed by a - al ulation on the omplete
losed-loop system.

The robust stability, nominal performan e and robust performan e plots of the
longitudinal system an be found in gure 31.8. Robust stability is a hieved,
indi ating that the ontroller stabilizes the omplete set of plant systems. With
the demanding set of requirements and the large plant perturbations, obtaining
nominal performan e over the omplete frequen y range was impossible. With
the weighting fun tions used, this was the best obtainable result. Nevertheless,
the hoi e of weighting fun tions is stringent enough to give good performan e
even with nominal performan e values above unity. The performan e does
not deteriorate mu h if the un ertainty and ight envelope perturbations are
introdu ed. This is indi ated by the robust performan e -plot whi h is only
slightly larger than the nominal performan e plot. An interesting feature of this
last plot is that be ause of the numeri al di ulty in the - al ulation with
the real valued perturbations used in this analysis, the upper and lower bounds
of  show a small gap. Similar results were a hieved in the lateral-dire tional
design, as an be seen in gure 31.9.

519
31.5.2 Linear time domain analysis
Figures 31.10 and 31.11 show the pit h and roll rate time responses to a step
input at t=1 se ond. To he k losed-loop system tra king of the ideal refer-
en e model, responses of both the ideal and the nominal system are plotted in
the same gure, together with the error or dieren e between them. Besides
the nominal system response, the response of the perturbed system with full
parametri un ertainties on the matrix elements is plotted also. This gives an
indi ation of the robustness of the resulting ontroller. Be ause the hoi e of
the dynami pressure variations explained in se tion 31.2, the nominal system
orresponds to a dynami pressure value of q = 6900.
Both gures show good tra king of the ideal response with little deterio-
ration when un ertainty perturbations are added. This gives an indi ation of
the robustness qualities of the ontrollers. Only at the onset of the step re-
sponse a small dieren e between the ideal and the a tual response exists, but
steady-state tra king is a hieved in both ases.

1.4 1.2

dotted: ideal pitch rate response

pilot
solid: nominal system
1.2 command
dashed: perturbed system 1
pilot
command
1
dotted: ideal roll rate response
0.8
solid: nominal system
0.8 dashed: perturbed system

0.6
0.6

0.4
0.4

0.2

q-error 0.2

0
p-error

0
-0.2

-0.4 -0.2
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
time [sec] time [sec]

Figure 31.10: Pit h ommand lin- Figure 31.11: Roll ommand linear
ear time response time response

31.6 Results of the Automated Evaluation

In the automated evaluation pro edure, the ontroller was subje ted to a se-
ries of nonlinear simulations to verify omplian e with the HIRM requirements.
The overall impression was that the ontroller showed good performan e and
robustness hara teristi s. The ontroller satised the riteria on erning dis-
turban e reje tion, stru tural oupling and limited noise and turbulen e ontrol
a tivity. Gibson riteria and departure prote tion also looked good. Only the
lateral hannel multivariable un ertainty riterion was not satised ompletely.
As an example two gures from the nonlinear simulation assessment manoeu-
vres are plotted.
In gure 31.12 the response is given for a pit h rate demand, at M = 0.3
and h = 5000 ft. This plot shows that the pit h response a hieved orresponds

520
to the ideal handling qualities referen e model. The airspeed varies be ause of
the in reased pit h attitude while the ontroller tries to ompensate for that
with an in rease in thrust.
In gure 31.13 the response is given for a roll rate demand, at M = 0.5 and
h = 15000 ft with parametri un ertainties and measurement errors present.
Despite the added un ertainties the roll rate response follows the ideal roll
model. The transients in the roll rate and sideslip responses of this gure are
aused by the 2 measurement error on . As a result of the instantaneous
appli ation of this error at t=0 se onds, the ontroller generates fast ontrol
surfa e dee tions to try to for e the measured sideslip ba k to zero.

31.7 Con lusions and Lessons Learned

Using the robust ontrol design methodology of H1 -, -synthesis, it was pos-
sible to design and analyse - ontrollers for the realisti ontrol augmentation
design problem stated in the HIRM design ben hmark. These ontrollers were
embedded in a nonlinear ontroller blo k ontaining dynami pressure on-
troller output s heduling and several nonlinear and limiting elements. This
resulting ontroller exhibits good performan e hara teristi s in terms of han-
dling qualities and departure prevention, while maintaining robustness against
model un ertainty and parameter variations.
The strong points of H1 -, -theory are:
 Multivariable design problems an be solved in a fully multivariable design
and analysis environment.
 The same framework is used for both synthesis and analysis of ontrollers
and losed-loop systems.
 Robustness of the resulting ontroller against model un ertainty, unmod-
elled dynami s and parameter variations an be in orporated into the design
pro ess via a well-dened set of plant models around a nominal model.
 Handling qualities requirements an easily be implemented into the design
pro ess with the use of an ideal referen e model.
 A powerful set of dedi ated software tools exists. The Matlab mutools tool-
box provides the designer with all fun tions, subroutines and data stru tures
needed for the engineering appli ation of H1 -, -synthesis.
 The design y le an be largely automated.
During this design, several points of onsideration and lessons learned on ern-
ing the use of H1 -, -synthesis were identied, some general and some spe i
to this design problem. These are:
 A thorough understanding of the theory is needed to ee tively design and
analyse ontrollers.
 A large and di ult part of the design y le onsists of hoosing and hang-
ing the weighting fun tions. This eort an be redu ed by experien e and
a good physi al understanding of the possibilities and limitations of the
systems at hand.
 The resulting ontrollers may have a large number of states needing good

521
 model redu tion te hniques.
 It is not always feasible to keep on tuning the weighting fun tions to for e
the -value below unity for all frequen ies. Often, design obje tives are too
ompli ated to be expressed exa tly as rst or even se ond order weighting
fun tions and good designs an be made with larger -values.
 It is however, sensible to sear h for a set of weighting fun tions that results
in a ontroller whi h satises robust stability. This way, use is made of
one of the strong points of H1 -, -theory, namely: to be able to guarantee
stability for a given set of plant models.
 The linear ontrollers designed with H1 -, -synthesis are not able to ope
with large nonlinearities in the model and the design requirements. A lim-
ited number of additional lters, fun tions and nonlinear blo ks an be
added to the ontroller ar hite ture to deal with these nonlinearities.
All in all, this design has learly demonstrated that H1 -, -synthesis has apa-
bilities whi h an be used to solve a realisti nonlinear air raft robust ontrol
law design problem. It is expe ted that H1 -,  -theory will prove to be a valu-
able tool to the European Air raft Manufa tures for the design and analysis of
Flight Control Systems.

522
 x 10
7 pw q az
10 10 5

5 5 10

(m/s^2)
(deg/s)

(deg/s)
0 0 15

5 5 20
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

Va alpha x 10
7 beta
101 25 5

100.5 0
20
(deg)

(deg)
(m/s)

100 5
15
99.5 10

99 10 15
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

q cmd dts x 10
6 dtd
6 0 6

4
4 5
(deg/s)

(deg)

(deg)

2
2 10
0

0 15 2
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

dcs x 10
7 dcd x 10
6 dr
2 5 6

4
0 0
(deg)

(deg)

(deg)

2
2 5
0

4 10 2
0 5 10 0 5 10 0 5 10
time (s) time (s) time (s)

thr1 thr2
100 100
(%)

(%)

50 50

0 0
0 5 10 0 5 10
time (s) time (s)

Assessment manoeuvre: pitch rate demand, Mach 0.3, 5000 ft

Figure 31.12: Responses to a pit h rate ommand at Ma h 0.3 altitude 5000 ft.

523
 pw q az
100 1 5

0
0

(m/s^2)
(deg/s)

(deg/s)
50 5
1
10

0 2 15
0 2 4 6 0 2 4 6 0 2 4 6
time (s) time (s) time (s)

Va alpha beta
162.5 6 4

162 4
2
(deg)

(deg)
(m/s)

161.5 2
0
161 0

160.5 2 2
0 2 4 6 0 2 4 6 0 2 4 6
time (s) time (s) time (s)

p cmd dts dtd

80 6 10

60 7 5
(deg/s)

(deg)

(deg)

40 8 0

20 9 5

0 10 10
0 2 4 6 0 2 4 6 0 2 4 6
time (s) time (s) time (s)

dcs dcd dr
1 4 5

0 2 0
(deg)

(deg)

(deg)

1 0 5

2 2 10

3 4 15
0 2 4 6 0 2 4 6 0 2 4 6
time (s) time (s) time (s)

thr1 thr2
10 10
(%)

(%)

5 5

0 0
0 2 4 6 0 2 4 6
time (s) time (s)

Assessm. man.: roll rate demand, Mach 0.5, 15000 ft, param. unc.

Figure 31.13: Responses to a roll rate ommand at Ma h 0.5 altitude 15000 ft.
with parametri un ertainties and measurement errors

524
32. Nonlinear Dynami Inversion and
LQ Te hniques

Batri e Es ande 1

Abstra t. The ontribution presented here for the HIRM design

hallenge onsists of an expli it model-following te hnique, using a
ombination of non-linear inversion (NDI) and LQ synthesis. This
te hnique allows us to over a wide ight envelope without the
problem of gain s heduling. It also aims at better re-usability of
the ontroller on dierent airframes. Here, it is tested on a set of
typi al manoeuvres at dierent operating points.

32.1 Introdu tion

The present hapter on erns the presentation of ONERA's parti ipation in
the HIRM design hallenge.
The ontrol law design whi h is here under study is based upon an expli it
model-following te hnique using a ombination of nonlinear dynami inversion
(NDI) and LQ synthesis, leading to a two-level ontroller stru ture:

 a feedforward ontroller (NDI) whi h addresses the problems of perfor-

man e;

 a feedba k ontroller (LQ) whi h addresses the problems of stability and

robustness.

The two methods are presented in the tutorial part of the book (see hap-
ters 9 and 5). Their implementation for the HIRM ase are des ribed in the
following se tions.

32.2 Controller Stru ture

The design of the feedforward ontroller uses a nonlinear model of the air raft
and the NDI theory to treat the nonlinear aerodynami terms as well as the
nonlinearities oming from kinemati and gyros opi ouplings.
The design of the feedba k ontroller uses the linear quadrati ontrol the-
ory to treat the deviations (whi h are assumed to be small) of the air raft
outputs xr from the model states xrm . The presen e of an integral term on the
1
O e National d'Etudes et de Re her hes Aerospatiales (ONERA), BA701, 13661 Salon
de Proven e Air

525
 ontrolled variables z makes the deviations tend asymptoti ally towards zero.
The overall stru ture of the ontroller is displayed in gure 32.1.

uffwd

zp + u
Feedforward xrm + +
Controller Kx
pilot zm + control
input +
vector

aircraft reduced xr
state vector
z
aircraft Kz
controlled vector +

Feedback
Controller

Figure 32.1: Overall ontroller stru ture

32.2.1 Feedforward ontroller

The formulation of the feedforward ontrol law is based upon the for e and
moment equations. The ontrolled variables are in the present ase:

 pit h rate q or angle-of-atta k  (depending on the value of ), dened

as q or  ,
 velo ity ve tor roll rate pW , dened as pW
 and sideslip demand  , dened as  .
Airspeed is also one of the ontrolled variables, but it is treated separately of
the others.
The general s heme of the feedforward ontroller is given in gure 32.2 The
equations for lift and lateral for es and for velo ity ve tor roll rate are written
as follows :

g
_ = q (p os  + r sin ) tan  + (nz + os os )= os 
V
g
_ = p sin  r os  + (ny + os sin ) (32.1)
V
g
pW = (p os  + r sin )= os  n (+ os os ) tan 
V z
where ny and nz are the load fa tors in stability axes, the ight path angle
and  the bank angle.

526
 pilot uffwd
inputs slow fast
dynamics dynamics inversion
.
pwc pc pd
qc qc .
qd
c rc .
c rd xm
u
m pm 1
m qm
m 1+Ts
rm
m
load m
Vm
factors NonLinear
ny Model
nz Aircraft
modelled (onboard
state model)
vector
xm

Figure 32.2: Feedforward ontroller stru ture

The moment equations for pit h, roll and yaw about the body axes are
given by:

Ixx p_ Ixz r_ = LA + LT + (Iyy Izz )qr + Ixz pq

Iyy q_ = MA + MT + (Izz Ixx )pr Ixz (p2 r2 ) (32.2)
Ixz p_ + Izz r_ = NA + NT + (Ixx Iyy )pq Ixz qr
where LA , MA , NA are the omponents of the aerodynami moment and LT ,
MT , NT the omponents of the moment due to the thrust.
The feedforward ontrol law is based on a dynami model of the air raft
whi h is formulated by the following dierential equations:

x_ = F (x; u)
z = h(x) (32.3)

In general (see equations 32.3) F and h are non-linear fun tions of state ve tor
x (and of input ve tor u for F ), z being the ve tor of ontrolled variables.
The following assumptions are made:

 it is assumed that the dynami s of angular rates are faster than those of
angle-of-atta k and sideslip,

 the inuen e of the ontrol surfa es dee tions on the translational dy-
nami s is negligible.

From these assumptions, the feedforward ontroller an thus be separated

into two steps: the rst one is related to the slow dynami s and the se ond one
is related to the fast dynami s.
These assumptions allow the repla ement of the ommanded values on 
and  with ommanded angular velo ities, by inversion of the for es equations

527
system 32.1. Then, the input ve tor u is derived expli itly through a rst-order
dierentiation of these angular velo ities, whi h allows a dire t appli ation of
the NDI te hnique. By linearizing the rotational dynami s about a trimmed
value u of the input, the nal system to be inverted (moment equations, with

_ = (p;_ q;_ r_)) is written as :

_ = f (xm ) + g(xm )(uffwd u) (32.4)

In this equation the state ve tor used to ompute the dynami s is a mod-
elled state ve tor xm , that is omputed by an on-board model. The on-board
model is here identi al to the real air raft model (nonlinear hirmex ode) and is
in luded in the feedforward system. It uses as its input ve tor the feedforward
ommands uffwd, and delivers as output the modelled state ve tor xm . The
linearization value u is a ltered value of the input ve tor uffwd.
The basi obje tive for the dynami inversion is to ompute the input ve tor
uffwd as a fun tion of the desired dynami s
_ d . This gives the feedforward
ontrol law:
uffwd = u + g 1[
_ d f (xm ) (32.5)

The dierent steps leading to equation 32.5 are detailed in the following
se tions.

Slow dynami s
As was mentioned above, it was hosen to perform the inversion in two su es-
sive steps orresponding to slow and fast dynami states. The system alled
slow dynami s treats the equations of for es and the kinemati equation for
velo ity roll rate, whi h were written in system 32.1. It is represented in g-
ure 32.3.
The purpose of this inversion is to deliver ommanded angular velo ities
(p ; q ; r ) from the pilot ommands whi h are here (pW ; q ;  ). It an be
noti ed that the ommanded pit h rate q is dire tly input from by the pilot.
As angular velo ities are the fast variables, they are used as inputs in the for e
equations. As _ and pW are dependent on in p and r, dening a model of
response for :
_d = ! ( d )
and inverting 32.1 gives:

p = p? os  os  r? sin 
r = p? os  sin  + r? os  (32.6)

with the variables:

g
p? = pW os  + sin  (nz + os os )
V
g
r? = ! (  ) (n + os sin )
V y
528
 m
m
m coord.
m transformation
Vm

ny
nz
(A/C) pilot
nya input:
qc
nza
pc
limiter
qc
rc

pc
slow rc
pilot dynamics
inputs:
pwc
c

Figure 32.3: Slow ontroller

where ! is the desired bandwidth hosen for  in a ordan e with handling

qualities requirements.
If we onsider an angle-of-atta k limiter, the resultant ommanded pit h
rate is taken as the minimum of the pilot ommand and of qmax dened below
in equation 32.7, based on the  dynami s equation and a bandwidth ! , i.e.:
g
qmax = ! (max ) + (p os  + r sin ) tan  (n + os os )= os 
V z
(32.7)

Fast dynami s
The system alled fast dynami s onsists of an inversion of the moment equa-
tions. The purpose of this inversion is to deliver a ontrol surfa es dee tions
ve tor uffwd for a given set of ommanded angular rates p ; q and r .
At this stage, the problem of ontrol allo ation is posed as the ase (like for
the HIRM model) where the air raft model has more inputs than the ontrolled
outputs, and when the use of the lassi al inputs does not allow su ient ontrol
power.
The angular a elerations, expressed as in equation 32.4, are shaped to the
desired dynami s:

_ d = (p_d ; q_d; r_d )t
whi h is of rst order in roll and yaw and of se ond order in pit h :

p_d = !p (p pd )

529
 Z
q_d = 2q !q (q qd ) + !q2 (q qd ))
r_d = !r (r rd ))
Here again, the parameters !p , !q , q , !r are hosen in order to fulll the
handling qualities requirements.
The ontrol law uffwd is then omputed via a generalized inverse of the
linearized moment ontrol matrix:

uffwd = u +
[g(xm )
+ (

_d f (xm )) (32.8)

where [:+ denotes the pseudo-inverse, and
a diagonal weighting matrix.

The pseudo-inverse orresponds to a minimization of the normalized ontrol
surfa es dee tions. Other solutions exist for ontrol allo ation, like daisy
haining, dire t allo ation, other generalized inverse, but they are still at the
resear h level [29 [69.

32.2.2 Robust feedba k ontroller

The purpose of the feedba k ontroller is to minimise the deviations between
the real air raft state and the referen e model (on-board model).
It is done by a linear quadrati method, known for its robustness, using
only a redu ed state ve tor xr , dened as a subset of the state ve tor x:
xr = (v; w; p; q; r)t
The referen e model states are also represented here by a redu ed ve tor xrm ,
whi h is:
xrm = (v; w; p; q; r)tm
The ve tor of ontrolled variables z orresponds to the angular variables om-
manded by the pilot:
z = (pW ; q;  )t
and it is ompared to the variables dedu ed from the referen e model ontained
in ve tor zm :
zm = (pW ; q;  )tm
The ontrol law ufb k is expressed by the following expression, in whi h
the integral term on the ommanded variables z ensures a good asymptoti
behaviour [169:

Z
ufb k = Kx (xr xrm ) + Kz (z zm ) (32.9)

where the gain matri es are obtained by minimization of a quadrati index:

Z 1
J= (x_r t Qxx_r + z tQz z + u_ t Qu u_ )dt
0
530
where
xr = xr xrm
z = z zm
u = u uffwd
The weighting matri es (Qx ; Qz and Qu ) of this index are tuned in order to
meet robustness requirements.

32.2.3 Autothrottle
The airspeed ommand is here treated separately from the other pilot om-
mands, by a spe i ontrol law alled autothrottle. The reason for this sepa-
rate treatment is for pra ti al simpli ity, and it is justied by the fa t that the
response time is onsiderably longer than for the other ommands.
It omputes the total thrust required as a fun tion of the ommanded air-
speed, by inversion of the drag equation. The ommanded airspeed is on-
verted into a ommanded a eleration by imposing a se ond order response.
The damping and frequen y of this pres ribed response have to be tuned a -
ording to performan e requirements. The autothrottle system is represented
in gure 32.4.

2 1
s
+ throttle
Vc + + + position
commanded m Inversion of
airspeed 2
cos cos thrust model
+
+
measured V
airspeed

computed sin (.) g
flight path angle trimmed
thrust

Figure 32.4: Autothrottle

32.2.4 Resulting ontroller omplexity and implementa-

tion issues
The ontroller presented here, based on the NDI-approa h for the rst level and
the LQ-method for the se ond level, is of relatively high omplexity and of high
dynami order. This hoi e has been made due to the in reasing omputing
power that will be available on modern air raft. The order ould be redu ed
but the resulting performan e would also be ae ted. One reason for this
omplexity is that the rst level outputs are ontrol surfa es dee tions.
The air raft model whi h is in luded in the feedforward an be simplied by
taking a simpler aerodynami model (e.g. analyti instead of based on look-up
tables). Here again, the required level of realism of this plant model is one of

531
the main hoi es of this approa h. A su iently realisti plant model permits
to a hieve optimal use of the ontrol power of the air raft.
The present model (HIRM) with its asso iated ontroller, was implemented
with su ess in a real-time omputing environment. This demonstrates that
despite its relative omplexity this method is not ex essively time- onsuming
for a real air raft.

32.3 Design Des ription

32.3.1 Design obje tives
The rst design obje tive is to get a general ontroller stru ture that an be
easily re-used for any air raft mode [71. Typi ally, on e the ontroller is de-
signed for one air raft, it should be possible to adapt it to another one with a
limited number of hanges. When the hanges on ern only the aerodynami
parameter tables of a given air raft (for example for the study of dierent aero-
dynami designs), the ontroller an be re-used without any modi ation. In
parti ular, the tuning of the desired dynami s remains valid independently of
the model onsidered.
In some ases ontrol saturations an o ur, e.g. in the ase of dierent
ontrol powers, and require some readjustment of the dynami parameters and
of the weighting matrix
whi h ontrols the distribution among the dierent
ontrol surfa es. Other solutions an be used, su h as pilot input ltering.
This possibility enabled us to take into a ount the su essive versions of
the air raft model used for this hallenge.
The se ond obje tive to be mentioned is to separate the tuning parameters
of the dynami s relative to the pilot inputs (performan e aspe ts) and to the
turbulen e inputs (robustness aspe ts).
Another of our obje tives is to have a ontroller whi h is able to treat and
evaluate the new ontrol te hnologies, like thrust ve toring, strakes, su tion,et .
In parti ular, with the NDI approa h, the problem of ontrol allo ation an be
studied in a quite general framework.

32.3.2 Design y le des ription

On e the ontroller stru ture exists, i.e. the inversion of the equations is for-
mulated, the design y le onsists of the following steps:

 tuning the bandwidth parameters in order to a hieve the performan e

requirements,

1. !q and q hara terize the ideal pit h response and are tuned by
onsidering the dropba k riterion;

2. !p is dire tly related to the roll mode time onstant by !p = 1=T ;

3. !r and ! are hosen by xing the dut h roll hara teristi s in fre-
quen y and damping; the riterion on the sideslip response an here
be he ked;

532
  adjusting the weighting matri es of the feedba k ontroller. This tuning
is done starting from an initial guess whi h is a diagonal matrix formed
with the inverses of the squares of the variables onsidered (for ea h
variable the value taken here is lose to the maximum value a eptable).
For example, the Qz matrix is initially taken as:

0 1 0 0 1
z12
Qz = B
 0 z1 02
C
A
0 0 1
2

z32

 performing robustness tests and if ne essary modifying the feedba k ma-

tri es.

32.3.3 The translation of HIRM design riteria into

method dependent obje tives
The NDI approa h presented here takes dire tly into a ount the design riteria
if they are expressed in terms of modes.
For example, the riterion on roll mode time onstant (0.4s) is dire tly
translated by: !p = 1=0:4.
The other performan e requirements in the time domain are not dire tly
introdu ed in the parameters of the desired dynami s. But, as there is a phys-
i al meaning for the adjustable parameters, their tuning is done by means of
time simulations, with elementary pilot inputs. The parameters are:

 !r ; ! ; !q ; q orresponding respe tively to the rst order dynami s in

yaw and sideslip, and to the se ond order dynami s in pit h,

 ! , asso iated to the dynami s in  in luded in the angle-of-atta k lim-

iter,

 !f ; f , orresponding to the se ond order dynami s for the air speed.

The robustness is he ked independantly and an ne essitate some iterations
in the design y le. The stability of the system being reliable, the tuning of
weighting matri es of the se ond level still requires some experien e.

32.3.4 Comments on the eort required

The hoi e of the parameters of the rst level dynami s are independant of the
airframe model. That means that a given set of parameters an be used with
the same validity provided that the air raft models of the real and on-board
are identi al and that the ontrols are not saturated.
This aspe t is one of the potential time-saving qualities of the inverse dy-
nami s approa h.
The adjustment of the se ond-level matri es an require iterations if the
robustness requirements are not satisfa tory.

533
32.4 Analysis of the Resulting Controller
The basis of this analysis are the design requirements listed in hapter 27.

32.4.1 Typi al time responses

The riteria on dropba k and sideslip response are plotted in gures 32.5 and
32.6. It is shown that the ontroller gives here satisfa tory results, although
the dropba k value is very small.

Dropback criterion

2.5

2
THETA (deg)

1.5

0.5

0
0 0.5 1 1.5 2 2.5
t (s)

Figure 32.5: Dropba k riterion

Typi al time responses are also useful in order to analyse the ontroller's
behaviour and eventually rene the tuning of the rst-level bandwidth param-
eters, and of the diagonal weighting matrix
between the dierent ontrol
surfa es. The simulations are performed with the rst level alone, without
the se ond level orre tion. In the following gures (32.7 to 32.11), the time
responses to simple step inputs are shown. In the ight ase 2001, strong aero-
dynami ouplings are present in the HIRM model. The results show that with
the NDI method, the air raft's response is identi al to the pres ribed dynami s,
in the ase where there are neither modelling errors nor saturated ommands.

534
 Step sideslip requirement

Upper boundary

1.2

Normalised sideslip angle 0.8

Lower boundary

0.6

0.4

0.2

0
0 1 2 3 4 5 6 7 8 9 10
Time(seconds)

Figure 32.6: Sideslip response riterion

32.4.2 Frequen y responses in the linear domain at nom-

inal ight ondition
Performan es
The ompatibility of the rst-level pres ribed dynami s (1st order for velo -
ity ve tor roll rate, 2nd order for pit h rate) with the Gibson riteria an be
he ked, as is shown in gures 32.12 and 32.13.

Robustness
The robustness of the feedba k ontroller is evaluated by plotting the open
loop Ni hols graphs. In that ase, no feedforward ontroller is onsidered. The
results are obtained with the standard Matlab linearisation tools (gures 32.14
and 32.15).

535
 VV roll demand at flight condition 4010
VV roll demand at flight condition 4010 10 10
100 1 15
0 0

DTright deg
DTleft deg
50 0 10
Pw deg/s

Q deg/s
10 10

R deg/s
0 1 5
20 20
50 2 0
30 30
100 3 5 0 2 4 6 8 10 0 2 4 6 8 10
0 5 10 0 5 10 0 5 10

10 10
150 10 168

DCright deg
100 166

DCleft deg
0 0
THETA deg
PHI deg

V m/s
50 164
5 10 10
0 162
10

50 15 160 20 20
0 5 10 0 5 10 0 5 10 0 2 4 6 8 10 0 2 4 6 8 10

1.5 1.5 5 1
20
1 1 0.5
0

SUCTION
10

DR deg
AY m/s2
AX m/s2

AZ m/s2

0.5 0.5
5 0 0
0 0 10
10 0.5
0.5 0.5 20

1 1 15 30 1
0 5 10 0 5 10 0 5 10 0 2 4 6 8 10 0 2 4 6 8 10

4600 8 0.5 2 2

6 0 1.5 1.5
THROTTL1

THROTTL2
ALPHA deg

BETA deg

4550
Hm

4 0.5 1 1
4500
2 1 0.5 0.5

4450 0 1.5 0 0
0 5 10 0 5 10 0 5 10 0 2 4 6 8 10 0 2 4 6 8 10

Figure 32.7: Roll rate response (Ma h 0.4, 10000ft)

Pitch demand at flight condition 2001, with AOA limiting

10 10
Pitch demand at flight condition 2001, with AOA limiting
0 0
0.1 8 0.5
DTright deg
DTleft deg

6 0.4 10 10
0.05
Pw deg/s

Q deg/s

R deg/s

4 0.3 20 20
2 0.2 30 30
0
0 0.1 0 5 10 15 0 5 10 15
0.05 2 0
0 5 10 15 0 5 10 15 0 5 10 15 10 10

0.3 40 75
DCright deg
DCleft deg

0 0
0.2 35
THETA deg

70
PHI deg

V m/s

0.1 30 10 10
65
0 25
20 20
0 5 10 15 0 5 10 15
0.1 20 60
0 5 10 15 0 5 10 15 0 5 10 15
1
12 0.6 8 20
0.5
SUCTION

10 9 10
DR deg

0.4
AY m/s2
AX m/s2

AZ m/s2

8 10 0 0

6 11 10
0.2 0.5
20
4 12
30 1
2 0 13 0 5 10 15 0 5 10 15
0 5 10 15 0 5 10 15 0 5 10 15
2 2
400 35 0.06

380 0.04 1.5 1.5

THROTTL1

THROTTL2

30
ALPHA deg

BETA deg

360 0.02
Hm

25 1 1
340 0
20 0.5 0.5
320 0.02

300 15 0.04 0 0
0 5 10 15 0 5 10 15 0 5 10 15 0 5 10 15 0 5 10 15

Figure 32.8: Pit h rate response (Ma h 0.2, 1000ft)

536
 Pitch demand at flight condition 4010
Pitch demand
x 10
7 at flight condition 4010 7
x 10 10 10
2 8 1
0 0
6

DTright deg
DTleft deg
0.5
1
Pw deg/s

Q deg/s
10 10

R deg/s
4
0
2 20 20
0
0.5
0 30 30
1 2 1 0 2 4 6 8 10 0 2 4 6 8 10
0 5 10 0 5 10 0 5 10
7
x 10 10 10
3 30 162

25

DCright deg
2 160

DCleft deg
0 0
THETA deg
PHI deg

20

V m/s
1 158
15 10 10
0 156
10

1 5 154 20 20
0 5 10 0 5 10 0 5 10 0 2 4 6 8 10 0 2 4 6 8 10
7
x 10 1
8 1 5
20
6 0.5 10 0.5

SUCTION
10

DR deg
AY m/s2
AX m/s2

AZ m/s2

4 0 15 0 0
10
2 0.5 20 0.5
20

0 1 25 30 1
0 5 10 0 5 10 0 5 10 0 2 4 6 8 10 0 2 4 6 8 10
8
x 10
4900 16 4 2 2

14 1.5 1.5
4800 2
THROTTL1

THROTTL2
ALPHA deg

BETA deg

12
Hm

4700 0 1 1
10
4600 2 0.5 0.5
8

4500 6 4 0 0
0 5 10 0 5 10 0 5 10 0 2 4 6 8 10 0 2 4 6 8 10

Figure 32.9: Pit h rate response (Ma h 0.4, 10000ft)

sideslip demand at flight condition 4010

10 10
sideslip demand at flight condition 4010 x 103
0.2 2 1
0 0
DTright deg
DTleft deg

1
0.1 0.5
10 10
Pw deg/s

Q deg/s

R deg/s

0
0 0 20 20
1
0.1 0.5 30 30
2

0.2 3 1 0 2 4 6 8 10 0 2 4 6 8 10
0 5 10 0 5 10 0 5 10
10 10
0.1 7.119 161.25

161.2
DCright deg
DCleft deg

0 7.118 0 0
THETA deg
PHI deg

161.15
V m/s

0.1 7.117
161.1 10 10
0.2 7.116
161.05

0.3 7.115 161 20 20

0 5 10 0 5 10 0 5 10 0 2 4 6 8 10 0 2 4 6 8 10

1.3 0.2 9.72 1

20
0.1 0.5
9.725
SUCTION

1.25 10
AY m/s2
AX m/s2

AZ m/s2

DR deg

0
9.73 0 0
0.1
1.2 10
9.735 0.5
0.2
20
1.15 0.3 9.74 30 1
0 5 10 0 5 10 0 5 10 0 2 4 6 8 10 0 2 4 6 8 10

4572 7.13 1.5 2 2

1
ALPHA deg

1.5 1.5
BETA deg

4571.95
THROTTL1

THROTTL2

7.12
Hm

0.5
1 1
4571.9
0
7.11 0.5 0.5
4571.85 0.5
0 5 10 0 5 10 0 5 10 0 0
0 2 4 6 8 10 0 2 4 6 8 10

Figure 32.10: Sideslip response (Ma h 0.4, 10000ft)

537
 speed demand at flight condition 2001
speed demand at flight condition 2001 10 10
0.02 0.04 0.1
0 0

DTright deg
DTleft deg
0.01 0.02
0.05 10 10
Pw deg/s

Q deg/s

R deg/s
0 0

0.01 0.02
20 20
0
0.02 0.04 30 30

0.03 0.06 0.05 0 20 40 60 0 20 40 60

0 20 40 60 0 20 40 60 0 20 40 60
10 10
0.1 21.46 140

DCright deg
DCleft deg
0 21.44 120 0 0
THETA deg
PHI deg

V m/s
0.1 21.42 100
10 10
0.2 21.4 80
20 20
0.3 21.38 60 0 20 40 60 0 20 40 60
0 20 40 60 0 20 40 60 0 20 40 60

1
10 0.15 6
20
8 7 0.5
0.1

SUCTION
10

DR deg
AY m/s2
AX m/s2

AZ m/s2

6 8
0 0
0.05
4 9 10
0 0.5
2 10 20

0 0.05 11 30 1
0 20 40 60 0 20 40 60 0 20 40 60 0 20 40 60 0 20 40 60

4500 30 0.01 2 2

25 1.5 1.5
0.005
THROTTL1

THROTTL2
ALPHA deg

BETA deg

4000
20
Hm

0 1 1
15
3500
0.005 0.5 0.5
10

3000 5 0.01 0 0
0 20 40 60 0 20 40 60 0 20 40 60 0 20 40 60 0 20 40 60

Figure 32.11: Speed response (Ma h 0.2, 1000ft)

25

20

15

10

5
OpenLoop Gain (db)

5 3

10 1

15 1 Sluggish Response

2
PIO prone

20 2 Good Response

3 Quick, Ratcheting

25
270 180 90
OpenLoop Phase (deg)

Figure 32.12: Linear roll Gibson riterion for 1st order roll dynami s

538
 20

15

10
OpenLoop Gain (db)

5 L2 L1 L2

L3
0
L2

L1

10

180 90
OpenLoop Phase (deg)

Figure 32.13: Linear pit h Gibson riterion for 2nd order pit h dynami s

539
 DTs loop open DTd loop open
40 40
0 db 0 db
0.25 db 0.25 db
0.5 db 0.5 db
OpenLoop Gain (db)

OpenLoop Gain (db)

20 1 db 1 db 20 1 db 1 db
3 db 3 db
6 db 3 db 6 db 3 db
0 6 db 0 6 db
12 db 12 db
20 20 db 20 20 db

40 40 db 40 40 db
360 270 180 90 0 360 270 180 90 0
OpenLoop Phase (deg) OpenLoop Phase (deg)
DCs loop open DCd loop open
40 40
0 db 0 db
0.25 db 0.25 db
0.5 db 0.5 db
OpenLoop Gain (db)

OpenLoop Gain (db)

20 1 db 1 db 20 1 db 1 db
3 db 3 db
6 db 3 db 6 db 3 db
0 6 db 0 6 db
12 db 12 db
20 20 db 20 20 db

40 40 db 40 40 db
360 270 180 90 0 360 270 180 90 0
OpenLoop Phase (deg) OpenLoop Phase (deg)
DR loop open
40
0 db
0.25 db
0.5 db
OpenLoop Gain (db)

20 1 db 1 db
3 db
6 db 3 db
0 6 db
12 db
20 20 db

40 40 db
360 270 180 90 0
OpenLoop Phase (deg)

Figure 32.14: Ni hols plots with su essive loops open

540
 DTs loop open DTd loop open
40 40
0 db 0 db
0.25 db 0.25 db
0.5 db 0.5 db
OpenLoop Gain (db)

OpenLoop Gain (db)

20 1 db 1 db 20 1 db 1 db
3 db 3 db
6 db 3 db 6 db 3 db
0 6 db 0 6 db
12 db 12 db
20 20 db 20 20 db

40 40 db 40 40 db
360 270 180 90 0 360 270 180 90 0
OpenLoop Phase (deg) OpenLoop Phase (deg)
DCs loop open DCd loop open
40 40
0 db 0 db
0.25 db 0.25 db
0.5 db 0.5 db
OpenLoop Gain (db)

OpenLoop Gain (db)

20 1 db 1 db 20 1 db 1 db
3 db 3 db
6 db 3 db 6 db 3 db
0 6 db 0 6 db
12 db 12 db
20 20 db 20 20 db

40 40 db 40 40 db
360 270 180 90 0 360 270 180 90 0
OpenLoop Phase (deg) OpenLoop Phase (deg)
DR loop open
40
0 db
0.25 db
0.5 db
OpenLoop Gain (db)

20 1 db 1 db
3 db
6 db 3 db
0 6 db
12 db
20 20 db

40 40 db
360 270 180 90 0
OpenLoop Phase (deg)

Figure 32.15: Ni hols plots with su essive loops open, with parametri un er-
tainties

541
32.5 Con lusions
The NDI/LQ approa h has been applied to the problem of robust ight ontrol
of a ghter air raft model, the HIRM. The main advantages that hara terize
the approa h are:

 its exibility in terms of its potential appli ation to a wide range of ve-
hi les,

 its rapid implementation and easy a omodation of desired handling qual-

ities,

 the good a ura y of the rst level, that allows, for a given level of tur-
bulen e, to have more manoeuvrability power,

 the independant tuning of performan e and robustness.

In parti ular, it ould be possible to use new ontrol means su h as nose

su tion without important hanges in the ontroller (a few hours of work). But
it also has some drawba ks:

 it does not work in ase of saturations,

 it does not take expli itely into a ount the robustness riteria,

 the feedforward ontroller is rather omplex.

542
33. The Robust Inverse Dynami s
Estimation Approah

Ewan Muir 1

Abstra t. A Robust Inverse Dynami s Estimation ontrol system

has been designed su essfully for the HIRM design problem. The
method has proved simple to use. Within the onstrained design
time, it has produ ed a ontroller, with a re ognisable stru ture
and physi ally meaningful gain matri es, whi h largely satises the
design riteria.

33.1 Introdu tion

Robust Inverse Dynami s Estimation (RIDE, see Chapter 10) is a design method
whi h has developed from two previous ontrol te hniques: the Salford Singular
Perturbation Method [115 and Pseudo-Derivative Feedba k [14. The method
an be viewed as a form of realisable Non-linear Dynami Inversion (NDI) due
to the use of a dynami inverse input whi h inverts the model with respe t
to the ontrolled outputs, thus providing output de oupling, and Proportional
plus Integral feedba k loops whi h ompensate for errors in the al ulation of
the dynami inverse input and assign the desired losed-loop response hara -
teristi s. The dynami inverse input is also similar to the equivalent ontrol
input used in Variable Stru ture Control.

33.2 Sele tion of the Controller Ar hite ture for

the HIRM Problem
33.2.1 Control law stru ture
As spe ied in the HIRM design do ument, the ontrol law inputs are pit h
rate, whi h in ludes positive limiting, velo ity ve tor roll rate and sideslip de-
mands. A speed ontroller is also implemented. Normal a eleration and neg-
ative angle of atta k limiting fun tions are missing due to la k of design time,
but implementation would be similar to the positive  limiting.
The ontrol law drives the tailplane, whi h is dee ted both symmetri ally
and dierentially, and the rudder. Symmetri anard dee tions ould also be

1
Defen e Resear h Agen y, Flight Dynami s and Simulation Department, Bedford, MK41
6AE, UK

543
in luded relatively easily to minimise trim drag and to provide additional pit h
ontrol when tailplane dee tion is limited. This has not been done in this
study due to time onstraints. Symmetri thrust demands have been used to
ontrol air speed.
The pilot inputs and ontrol surfa e demands listed above are built into the
omplete ontrol law stru ture shown in Fig. 33.1, the omponents of whi h
are des ribed in further detail in se tions 33.2.2 to 33.2.6 below.

- thrust
VT K thrust 0 demand table look-up
dem PT (CB)-1
T demand corrected thrust to
+ throttle
for altitude
throttle demand

feedforward symm.
h
0 canard
ax, az ax, az

- diff.
0
q canard
dem Pitch rate -1 K
T V Aircraft
demand / qdem + VT
AoA Mach 2(s+20)
2
symm. Model
limit
limiter (s+40)2 tail
+ 2 p, q, r
Mux

diff.

Demux
K 2(s+20)
I + + (s+40)2 tail
p p + - + + -
vvdem dem
Velocity Mach rudder
K ,
dem vector rdem P Mach
roll /
demand
, u^ di

Ku
di u, v, w Calculate u,v and w
based on VT, and

Core RIDE controller Mach

Figure 33.1: Stru ture of RIDE ontrol law for HIRM

33.2.2 Core RIDE ontroller

The ore RIDE ontroller, identied in Fig. 33.1 above, onsists of the follow-
ing omponents. The dynami inverse input u^di de ouples the outputs, i.e. the
body axis rates, from ea h other and the other air raft states by using moment
an ellation. The proportional loop with gain KP a ts to stabilise the air raft
and is added in this way to provide pseudo-derivative feedba k. The integral
path on the ommands with gain KI provides robustness to errors in the esti-
mate of the dynami inverse input u^di . The PI ontroller enables the designer
to spe ify the losed-loop dynami s whi h are se ond order of the form

y = (s2 I + 2Zd
n s +
n 2 ) 1
n 2 y (33.1)

The gain matrix al ulations are elementary:

KP = (CB ) 1 2Zd
n (33.2)

KI = (CB ) 1
n 2 (33.3)

544
 KV = (CB ) 1 M (33.4)

Kudi = (CB ) 1 CA (33.5)

where Zd,
n and M are all diagonal. Zd denes the damping and
n the
frequen y of the losed-loop system response, while M spe ies the overall
gain of the feedforward input (the time onstant of the feedforward washout is
set by T ).

33.2.3 Feedforward
The feedforward is a washout whi h provides an additional degree of freedom
in tuning the initial air raft response to a pilot demand by adding the following
input to the integral path

uff = TKsV+sI y
where (33.6)
KV = (CB ) 1 M
As both M and T
1 are diagonal matri es, ea h of the demands an be viewed
separately with the feedforward inputs being distributed to the ontrol surfa es
by the inverse motivator ee tiveness matrix (CB ) 1 .

33.2.4 Pit h rate demand / AoA limiter

The pit h rate demand system is provided by al ulating a body axis pit h rate
using equation 33.7.

qdem = os qdem + sin pvvdem (33.7)

To provide an angle of atta k limiting fun tion, the pit h rate required to y
the air raft at the limit AoA is al ulated from equation 33.8.

anz a
qdem = _ dem + (p tan  ) os  + (r tan  + nx ) sin  (33.8)
VA os  VA os 
where _ dem is al ulated from equation 33.9

_ dem = (limit )bw (33.9)

where bw is the bandwidth of the AoA ontrol loop.

The pit h rate demand used to al ulate the ontrol surfa e dee tions is
the minimum of the onventional pit h rate demand, al ulated from equation
33.7 and the pit h rate required to limit the angle of atta k to the maximum
value, al ulated from equation 33.8. Note that to date, this has only been
developed to limit the positive angle of atta k. A similar s heme ould be
introdu ed for limiting on the negative AoA limit.

545
33.2.5 Velo ity ve tor roll / sideslip demand
Body axis roll and yaw rates are al ulated from the demanded velo ity ve tor
roll rate and sideslip demands using equations 33.10 to 33.13. The body axis
roll rate demand of the RIDE ontroller is given by

pdem = os  os pvvdem + sin (_dem rvv ) (33.10)

The body axis yaw rate demand of the RIDE ontroller is given by

rdem = sin  os pvvdem os (_dem rvv ) (33.11)

where the velo ity ve tor yaw rate, rvv , is al ulated from

rvv = os r sin p (33.12)

and the demanded rate of hange of sideslip _ dem is given by

_dem = (dem  )bw (33.13)

where bw is the bandwidth of the sideslip ontrol loop.

33.2.6 Air speed ontroller

The speed demand is me hanised as a simple error driven proportional on-
troller. The error demands a thrust whi h is translated ba k into a throttle
setting.

33.2.7 Filters
During the design, a lter was in luded on the stabilator demand, both sym-
4(s+20) 2
metri and dierential, to improve the phase margins. The lter used is
(s+40)2

. The hoi e of this lter is in no sense optimal and was hosen by the rapid
trial of lters of the same order but with varying time onstants and observ-
ing their ee t on Ni hols plots. The results from se tion 33.5 indi ate that
this solution is adequate and satises the majority of the riteria. However, it
does in rease the high frequen y ontrol law gain from pit h rate to tailplane.
In retrospe t, a redu tion in the gains
n may have produ ed adequate gain
margins without re ourse to these lters.

33.2.8 Measurement signals

As the ontroller regulates prin ipally the body axis rates, the primary feed-
ba k signals are p, q and r. Air speed is obviously required for the speed
ontroller. The body axis rates are also required in the al ulation of the dy-
nami inverse input, together with body axis velo ities. These velo ities have
been supplied based on the total air speed resolved into body axes using 
and . Measurements of  and  are also used to al ulate the body axis roll
and yaw rate demands ne essary to provide the velo ity ve tor roll and sideslip

546
demand s hemes. The  limiter needs AoA, air speed and longitudinal and
normal a eleration feedba k.
Ma h number is used for s heduling for onvenien e as it is readily available
as an output from the model. Ideally, dynami pressure would be used as the
s heduling variable.

33.2.9 Resulting ontroller omplexity and implementa-

tion issues
The ontrol law diagram in Fig. 33.1 shows that the omplete RIDE ontroller
(CB ) 1 and
is simple. For the body axis rate demand system, only the matri es
1
((CB ) CA) need to be implemented as look-up tables s heduled with Ma h
number. In this ase, the 5 design points are those for whi h the linearised
models are available, i.e. M0.2, 1k ft; M0.24, 20k ft; M0.3, 5k ft; M0.4, 10k
ft; M0.5, 15k ft. The gain matrix equations 33.2 to 33.5 are simple and the
omputational load involved in both generating and running a RIDE ontroller
is therefore minimal. The dynami elements are restri ted to the six integrators
and the two se ond order lters used to improve the phase margins.
The speed demand system only has one s alar gain with three one-dimensional
look-up tables.
The rest of the ontroller, whi h translates pilot demands into body axis
rates and provides limiting fun tions, will be of a similar order of omplexity
irrespe tive of the design method used.

33.2.10 Further extensions

The urrent ontroller omits the logi required to handle ontrol surfa e position
and rate limiting. This logi is simple and would not impose a signi ant
omputational load. Position limiting is handled by holding the integrators
on the integral and feedforward portions of the ontroller and rate limiting
is handled by initialising the value of the integrator to give the maximum
in remental ontrol surfa e dee tion possible given the rate limit.

33.3 Translation of Design Criteria into Method

Dependent Obje tives
The only riteria whi h map onto the RIDE design method are those for band-
width whi h are dire tly governed by the sele tion of the
n matri es. The
matrix of damping, Zd , has been set to unity. In addition, the sele tion of the
M and T matri es also ae ts the rise time initially. They have been used
to tune the pit h rate response to provide the required dropba k and and to
ensure that the roll and yaw rate responses have a eptable rise times.
The method ontains no me hanism for expli itly guaranteeing robustness
and thus any robustness analysis would need to be done separately using an
additional tool.

547
33.4 Des ription of the Design Cy le
33.4.1 Extra t CB and CA information at sele ted trim
points
Look-up tables are onstru ted with an adequate spread a ross the ight enve-
lope. Sele tion of the s heduling points relies on engineering judgement. Past
experien e [176 shows that s heduling with Ma h number and/or dynami
pressure is su ient. In the HIRM design, the A and B matri es were taken
from the 5 linearised models identied in se tion 2.6.

33.4.2 Initial gain sele tion

The natural frequen y,
n , and damping, Zd , are sele ted rst. For the HIRM
example, damping in ea h of the hannels was set to 1, i.e. Zd = I3x3 . The
frequen ies
n were sele ted to give approximately the orre t bandwidths of
response. The pit h frequen y was hosen to be 5 r/s and the roll frequen y
was set to 4 r/s. A yaw frequen y of 4 r/s was also sele ted to ensure that
onsistent bandwidths are used for the velo ity ve tor rolls where a mix of
body axis roll rate and yaw rate is required.
One limitation whi h needs to be onsidered is that as the gain matri es
KP , KI and KV are fun tions of (CB ) 1 , as shown in equations 33.2 to 33.4,
the demanded response bandwidth, spe ied by
n , will need to be redu ed for
ight onditions where the motivator ee tiveness is greatly redu ed, in order
to avoid una eptably large gains.

33.4.3 Control law tuning

Having sele ted the feedba k gain hara teristi s and he ked the initial re-
sponse of the basi losed-loop system, the feedforward gains, T 1 and M,
are then hosen to give the orre t amount of dropba k in pit h and a suit-
ably rapid onset of response in roll and yaw. T 1 modies the rate at whi h
the feedforward input washes out and M hanges the gain on the feedforward
input. The nal values arrived at were

2 3 2 3
3 0 0 3 0 0
T 1 = 4 0 10 0 5 and M = 4 0 10 0 5
0 0 5 0 0 1
for the p, q and r demands.
The proportional gain on the speed ontroller was hosen by in reasing
its value until a response with an adequate rise time was a hieved with no
overshoot in speed. A value of 0.35 was eventually sele ted. Similarly, the
optimum bandwidths on the angle of atta k demand, bw , (set to 5 rad/s) and
on the sideslip demand bw (set to 2 rad/s) were sele ted by in reasing the
bandwidths to provide the maximum performan e with adequate damping.

548
 Trade os an be made between the magnitude of the inner loop gains,
spe ied by
n (whi h will ae t how lose the response is to that predi ted
theoreti ally but whi h will be limited due to stability onsiderations), against
the feedforward gains (whi h boost the initial response but may lead to an
overshoot). For the HIRM design, tuning was performed by sele ting values of
frequen y and damping whi h would give an adequate speed of response but
not introdu e instabilities when the sensors and a tuators were in luded. If the
initial response speed was inadequate, the feedforward gains were adjusted to
give the required speed of response and overshoot.

33.4.4 General omments on synthesis pro edure

The ore HIRM ontroller an be produ ed simply; the gain matri es an easily
be al ulated by hand. De oupling is taken are of automati ally. Tuning
of the losed-loop responses is performed based on the re ognisable ee t of
ea h of the parameters. On e the basi RIDE ontrol law has been developed,
produ tion of the rest of the ontroller is method independent. The me hanisms
for handling ontrol surfa e position and rate limiting would be straightforward
to implement.

33.5 Analysis of the Resulting Controller in Terms

of the Applied Methodology
33.5.1 Information on time and frequen y response plots
The time responses were produ ed using the omplete ontroller applied to the
non-linear model with the full, detailed sensor representation. The frequen y
responses were produ ed by linearising the non-linear losed-loop system and
then analysing it.
It should be noted that all of the time responses are plotted against time
in se onds. In the time responses, the pit h rate, velo ity ve tor roll rate and
sideslip demands are plotted using dashed lines. The symmetri al tailplane
dee tion is plotted as a solid line in the ontrol dee tions subplot, the dier-
ential tailplane dee tion is plotted as a dashed line and the rudder dee tion
as a dotted line. Frequen ies marked on the frequen y responses are in rad/s.
In Figs. 33.9 to 33.11 and 33.15 to 33.17, the frequen y responses with
perturbations are plotted as dot-dashed lines to allow the ee t of the pertur-
bations to be observed.

33.5.2 Responses using non-linear air raft model

Small amplitude non-linear time responses (Ma h 0.4, 10k ft)
The small amplitude pit h rate response shown in Fig. 33.2 shows that the
air raft behaves in a broadly a eptable manner although a slight os illation is

549
noti eable on the overshoot on the pit h rate response. The os illation is due
to the se ond order lter used to improve the phase margin.

9
x 10
1.5 10

vv roll rate (deg/s)

pitch rate (deg/s)
1
5
0.5
0
0

0.5 5
0 2 4 6 0 2 4 6
10
x 10
10.5 10

bank angle (deg)

10
AoA (deg)

5
9.5
0
9

8.5 5
0 2 4 6 0 2 4 6
9
x 10
0.405 2
Mach number

0.4 1

sideslip (deg)
0.395 0

0.39 1

0.385 2
0 2 4 6 0 2 4 6
4
x 10
1.006 5
control deflections (deg)

1.004
altitude (ft)

0
1.002
5
1

0.998 10
0 2 4 6 0 2 4 6
Time (secs) Time (secs)

Figure 33.2: Small amplitude pit h rate response (Ma h 0.4, 10k ft)

Despite the os illations, the attitude response, shown in Fig. 33.3, is

smooth. Hen e, this os illation may be imper eptible to the pilot. The peak
stabilator dee tion is a hieved in 0.11s and so the peak pit h a eleration
should also be rea hed in a similar time s ale, whi h meets the 0.15s riterion
spe ied in the design do ument.
The pit h attitude response in Fig. 33.3 shows that the dropba k is nearly
zero and satises the relevant design riterion.
The small amplitude roll rate response plotted in Fig. 33.4 shows that the
ontrol system satises the roll ommand with a time onstant of approximately
0.35s as spe ied. There is little oupling into either pit h rate or sideslip and
the ontrol a tivity is small.
The air raft response to a sideslip demand, shown in Fig. 33.5, is good. The
sideslip response lies within the boundaries of Fig. 27.16 of hapter 27. The
manoeuvre is a ompanied by very little roll and no pit h rate. The ontrol
a tivity is also moderate.

Large amplitude non-linear time responses (Ma h 0.4, 10k ft)

The large amplitude pit h rate response shown in Fig. 33.6 is well behaved
with the angle of atta k being limited at 30o. The stabilator dee tions used
to produ e the manoeuvre are a eptable although some os illations do appear

550
 12

11.5

11

pitch attitude (deg)

10.5

10

9.5

8.5
0 1 2 3 4 5 6
Time (sec)

Figure 33.3: Pit h attitude response to al ulate dropba k (Ma h 0.4, 10k ft)

0.4 1.5
vv roll rate (deg/s)
pitch rate (deg/s)

0.3
1
0.2
0.5
0.1
0
0

0.1 0.5
0 2 4 6 0 2 4 6

8.9 4
bank angle (deg)

8.85
3
AoA (deg)

8.8
2
8.75
1
8.7

8.65 0
0 2 4 6 0 2 4 6

0.402 0.15
Mach number

0.4 0.1
sideslip (deg)

0.398 0.05

0.396 0

0.394 0.05
0 2 4 6 0 2 4 6

10000.5 2
control deflections (deg)

10000 0
altitude (ft)

9999.5 2

9999 4

9998.5 6

9998 8
0 2 4 6 0 2 4 6
Time (secs) Time (secs)

Figure 33.4: Small amplitude vv roll rate response (Ma h 0.4, 10k ft)

551
 0.4 0.4

vv roll rate (deg/s)

pitch rate (deg/s)
0.3
0.2
0.2
0
0.1
0.2
0

0.1 0.4
0 2 4 6 0 2 4 6

8.9 0.4

bank angle (deg)

8.85
0.2

AoA (deg)
8.8
0
8.75
0.2
8.7

8.65 0.4
0 2 4 6 0 2 4 6

0.402 1.5
Mach number

0.4 1

sideslip (deg)
0.398 0.5

0.396 0

0.394 0.5
0 2 4 6 0 2 4 6

10000.5 2

control deflections (deg)

10000 0
altitude (ft)

9999.5 2

9999 4

9998.5 6

9998 8
0 2 4 6 0 2 4 6

Figure 33.5: Response to small amplitude sideslip demand (Ma h 0.4, 10k ft)

after 7 se onds. The reason for the de rease in stability is due entirely to the use
of Ma h number instead of dynami pressure for s heduling. For this response,
the Ma h number has redu ed to around M0.25, hen e the gain look-up tables
will be using gains based on those for the M0.24, 20k ft where the air density
is approximately 0.6. However the air raft is only at 10k ft (air density 0.9)
and so the gains will be 50% too large, leading to the redu tion in stability.
S heduling with dynami pressure would learly have been preferable. The
oupling into roll and sideslip is minimal.
The large amplitude roll response in Fig. 33.7 shows that the air raft is
well behaved during the 360o roll. However there is a moderate amount of
oupling into pit h whi h is unexpe ted. This is due to gravity ee ts whi h
are not being a ounted for. Also, the maximum sideslip ex ursion during
the roll is 3o whi h is in ex ess of that permitted. S ope for redu ing this is
limited however, as an in rease in the bandwidth of the demand loop leads to
instability.
The large amplitude sideslip demand produ es little oupling into roll and
negligible oupling into pit h. The demand is satised with a response similar
to the small amplitude one shown in Fig. 33.5.

Frequen y responses from linearised non-linear model (Ma h 0.4, 10k

ft)
Ea h of the frequen y responses shown in Figs. 33.9 to 33.11 avoids the ex lu-
sion zone. It an also be seen that the perturbations have little ee t.

552
 20 1

vv roll rate (deg/s)

pitch rate (deg/s)
0.5
10
0

0.5
0
1

10 1.5
0 2 4 6 8 0 2 4 6 8

30 1

bank angle (deg)

25
0.5

AoA (deg)
20
0
15
0.5
10

5 1
0 2 4 6 8 0 2 4 6 8

0.45 0.5

0.4
Mach number

sideslip (deg)
0
0.35

0.3
0.5
0.25

0.2 1
0 2 4 6 8 0 2 4 6 8
4
x 10
1.06 10
control deflections (deg)
5
1.04
altitude (ft)

0
1.02
5
1
10

0.98 15
0 2 4 6 8 0 2 4 6 8

Figure 33.6: Response to large amplitude pit h rate demands with  limiting
(Ma h 0.4, 10k ft)

4 80
vv roll rate (deg/s)
pitch rate (deg/s)

2 60

0 40

2 20

4 0

6 20
0 2 4 6 8 0 2 4 6 8

10 400
bank angle (deg)

300
AoA (deg)

5
200
0
100

5 0
0 2 4 6 8 0 2 4 6 8

0.43 4
Mach number

0.42 2
sideslip (deg)

0.41 0

0.4 2

0.39 4
0 2 4 6 8 0 2 4 6 8

10200 10
control deflections (deg)

10000
altitude (ft)

0
9800
10
9600

9400 20
0 2 4 6 8 0 2 4 6 8

Figure 33.7: Large amplitude 360o roll at 70o/s (Ma h 0.4, 10k ft)

553
 0.4 4

vv roll rate (deg/s)

pitch rate (deg/s)
0.3
2
0.2
0
0.1
2
0

0.1 4
0 2 4 6 8 0 2 4 6 8

9.1 1

bank angle (deg)

9
0

AoA (deg)
8.9
1
8.8
2
8.7

8.6 3
0 2 4 6 8 0 2 4 6 8

0.405 15
Mach number

10

sideslip (deg)
0.4
5
0.395
0

0.39 5
0 2 4 6 8 0 2 4 6 8
4
x 10
1.0004 15

control deflections (deg)

1.0003 10
altitude (ft)

1.0002 5

1.0001 0

1 5

0.9999 10
0 2 4 6 8 0 2 4 6 8

Figure 33.8: Response to large amplitude sideslip demand (Ma h 0.4, 10k ft)

Small amplitude non-linear time responses (Ma h 0.24, 20k ft)

From the small amplitude pit h rate response in Fig. 33.12, it an be seen that
there is a small de rease in the damping of the system but the demands are still
tra ked well and there is little oupling into roll or sideslip. The low damping
apparent in the rst se ond of the responses is due to impre ise initialisation
and so should be disregarded, as it is not apparent elsewhere in the responses.

The small amplitude roll rate response in Fig. 33.13 shows similar hara -
teristi s to the previous plot in Fig. 33.4 at M0.4 although there is an in rease
in the amount of oupling into both pit h rate and sideslip.

Comparing the sideslip responses in Fig. 33.14 with those in Fig. 33.5,
shows that, despite the hange in ight ondition, the air raft exhibits a sim-
ilar response to the demand. However it an be seen that de rease in the
ontrol surfa e ee tiveness at the higher angle of atta k and altitude results
in signi antly greater ontrol surfa e dee tions.

Large amplitude demands have not been made for the M0.24 ase as the
ontrol law does not have onditioning of the integrators, in the event of ontrol
surfa e position and rate limiting, and large demands invariably lead to both
types of saturation on at least one of the surfa es.

554
 40

0.2382
30
0.5672
1.351

20

3.218
10

0 7.663

18.25
10
43.47

20

30

40
350 300 250 200 150 100 50 0

Figure 33.9: Symmetri tailplane freq. resp.

0.1
40

0.2382
30
0.5672

20
1.351

10
3.218

0
7.663

18.25
10

43.47
20

30

40
350 300 250 200 150 100 50 0

Figure 33.10: Dierential tailplane freq. resp.

Frequen y responses from linearised non-linear model (Ma h 0.24,

20k ft)

The symmetri and dierential tailplane frequen y responses both avoid the
ex lusion zone, although only marginally in the symmetri tailplane ase. The
rudder response does violate the zone. Rudder was the one surfa e not to
have any additional ltering added to improve the phase margin. Su h ltering
should be able to re tify this. As for the Ma h 0.4 ase, it an be seen that the
perturbations have little ee t, although they do ae t the response at higher
frequen ies than before.

555
 40
0.1

30 0.2382

0.5672
20
1.351

10 3.218

0 7.663

18.25
10

20
43.47

30

40
350 300 250 200 150 100 50 0

Figure 33.11: Rudder frequen y response

0.5 0.2
vv roll rate (deg/s)
pitch rate (deg/s)

0 0.1

0.5 0

1 0.1

1.5 0.2
0 2 4 6 0 2 4 6

30 0.2
bank angle (deg)

29 0.1
AoA (deg)

28 0

27 0.1

26 0.2
0 2 4 6 0 2 4 6

0.255 0.2
Mach number

0.25
sideslip (deg)

0.1
0.245
0
0.24

0.235 0.1
0 2 4 6 0 2 4 6
4
x 10
2.0005 10
control deflections (deg)

5
2
altitude (ft)

0
1.9995
5
1.999
10

1.9985 15
0 2 4 6 0 2 4 6

Figure 33.12: Response to small amplitude pit h rate demand (Ma h 0.24, 20k
ft)

556
 0.2 1.5

vv roll rate (deg/s)

pitch rate (deg/s)
0
1
0.2
0.5
0.4
0
0.6

0.8 0.5
0 2 4 6 0 2 4 6

30 4

bank angle (deg)

3
29.5

AoA (deg)
2
29
1
28.5
0

28 1
0 2 4 6 0 2 4 6

0.25 0.4

0.3
Mach number

sideslip (deg)
0.245
0.2

0.1
0.24
0

0.235 0.1
0 2 4 6 0 2 4 6
4
x 10
2.0005 10

control deflections (deg) 5

altitude (ft)

2
0
1.9995
5

1.999 10
0 2 4 6 0 2 4 6

Figure 33.13: Response to small amplitude vv roll rate demand (Ma h 0.24,
20k ft)

0.2 0.4
vv roll rate (deg/s)
pitch rate (deg/s)

0
0.2
0.2
0
0.4
0.2
0.6

0.8 0.4
0 2 4 6 0 2 4 6

30 0.1
bank angle (deg)

0
29.5
AoA (deg)

0.1
29
0.2
28.5
0.3

28 0.4
0 2 4 6 0 2 4 6

0.25 1.5
Mach number

1
sideslip (deg)

0.245
0.5
0.24
0

0.235 0.5
0 2 4 6 0 2 4 6
4
x 10
2.0005 15
control deflections (deg)

10
altitude (ft)

2
5

0
1.9995
5

1.999 10
0 2 4 6 0 2 4 6

Figure 33.14: Response to small amplitude sideslip demand (Ma h 0.24, 20k
ft)

557
 40 0.1
0.2382
0.5672

30

1.351
20

10 3.218

0 7.663

18.25
10
43.47

20

30

40
350 300 250 200 150 100 50 0

Figure 33.15: Symmetri tailplane freq. resp.

40

30
0.5672
0.2382
1.351 0.1
20

10 3.218

0 7.663

18.25
10
43.47
20

30

40
350 300 250 200 150 100 50 0

Figure 33.16: Dierential tailplane freq. resp.

40

30 0.1
0.2382
0.5672
20
1.351

10
3.218

0 7.663

18.25
10

20
43.47

30

40
350 300 250 200 150 100 50 0

Figure 33.17: Rudder frequen y response

558
33.5.3 Responses from sti k to attitudes (Ma h 0.4, 10k
ft)
Gibson riteria
Figures 33.18 and 33.19 show that, given a suitable sti k s aling, the frequen y
response from sti k to attitude satises the Gibson riteria for the Ma h 0.4,
10k ft ase.

Gain and phase rate riteria

When the gain at 180o phase is examined, based on the frequen y responses
from sti k to attitude, it is apparent that the design does not satisfy the <-16
dB riterion as the gains for the system are as follows:

Pit h: -13.8dB at 180o phase

Roll: -10.1dB at 180o phase

However, the phase rate for ea h system is a eptable and both systems are
level 1* as the frequen y at 180o phase and the phase rate are:
Pit h: 2.32Hz at 180o phase with a phase rate of 44:6o/Hz
Roll: 1.02Hz at 180o phase with a phase rate of 310/Hz

20

15

10
Gain (db)

5
L1

5
L1

10

180 160 140 120 100 80 60 40

Phase (deg)

Figure 33.18: Sti k to pit h attitude response

559
 25

20

15

10

5
Gain (db)

0
Sluggish
5 PIO
Oscillation
10

15
Good
20

25
350 300 250 200 150 100 50 0
Phase (deg)

Figure 33.19: Sti k to roll angle response

560
33.5.4 Stability analysis (M0.4, 10k ft and M0.24, 20k ft)
The frequen y responses of Figs. 33.20 to 33.22 have been produ ed as follows.
For Figs. 33.20 and 33.22, a gain has been introdu ed on the tailplane and
rudder demand paths. This gain has been set to 1.68 (4.5dB) on ea h of
the tailplane and rudder demand paths simultaneously. A time response has
then been run with a small impulse of 0.1 deg/s amplitude and 0.1 se onds
duration, being made at 0.5 se onds applied to the pit h rate, vv roll and
sideslip demands. The time response has then been he ked to ensure that the
system has remained stable.
For Fig. 33.21, the gain was set to 1.33 (2.5dB). The natural frequen y
of the pit h rate response in Fig. 33.20 was measured, pit h rate response
having been hosen as this was learly the least well damped and most prone
to instability. Based on the natural frequen y (3.6Hz), a time delay whi h
would give 30o phase lag at this frequen y (23 mse ) was then inserted on ea h
of the tailplane and rudder demand loops simultaneously in addition to the
2.5dB in rease in gain.
Figs. 33.20 and 33.21 show that at M0.4, the losed-loop system remains
stable despite the simultaneous gain and phase additions.

0.5

0.4

0.3
Body axis rates (deg/s)

0.2

0.1

0.1

0.2

0.3
0 0.5 1 1.5 2 2.5 3 3.5 4

Figure 33.20: Additional 4.5dB gain at the M0.4, 10k ft ight ondition

For the Ma h 0.24 ase, Fig. 33.22 shows that the system remains just stable
with an additional gain of 4.5dB. The stability riterion is therefore met. It
was impossible to he k whether the system remained either stable or neutrally
stable with simultaneous gain and phase osets as the ontrol surfa es started
to rate limit. As integrator onditioning in the event of ontrol surfa e rate
limiting has not yet been added this led to instability and han e masking the
ee t of the additional gain and phase.

33.5.5 Response of speed ontroller

Fig. 33.23 shows the response to a large amplitude demand on speed. It an be
seen that the engine thrust in reases to the maximum level to provide a rapid

561
 0.8

0.6

Body axis rates (deg/s) 0.4

0.2

0.2

0.4

0.6
0 0.5 1 1.5 2 2.5 3 3.5 4

Figure 33.21: Additional 2.5dB gain and 30o phase at the M0.4, 10k ft ight
ondition

0.4

0.2

0
Body axis rates (deg/s)

0.2

0.4

0.6

0.8

1.2
0 0.5 1 1.5 2 2.5 3 3.5 4

Figure 33.22: Additional 4.5dB gain at the M0.24, 20k ft ight ondition

562
a eleration. A small steady-state oset, from the demanded 90m/s air speed,
is observed. In pra ti e, this should not be a problem.

90 380

85
360

air speed (m/s)

altitude (m)
80
340
75
320
70

65 300
0 5 10 0 5 10

4
x 10
8 22

20
6

AoA (deg)
thrust (N)

18
4
16
2
14

0 12
0 5 10 0 5 10

Figure 33.23: Response to large amplitude speed hange demand

33.6 Results of the Automated Evaluation Pro-

edure
The results from the omplete set of evaluations are as follows (see [177 for
the meaning of test ase numbers).
- Tests 11, 13 and 14 passed, agreeing with the results from Figs. 33.9 to
33.11 and 33.15 to 33.17.
- Test 12 failed, disagreeing with the results shown in Figs. 33.20 to 33.22.
- In test 21, the following values of RMS disturban e amplitude were re orded
due to moderate turbulen e:

0.5002/s in roll rate

0.0236/s in pit h rate
0.9901/s in yaw rate
0.0938g in normal a eleration.

- The responses for test 22 were impossible to interpret but this riterion is
overed at one ight ondition by the a eptable responses in Figs. 33.18 and
33.19.
- Some of the responses to the assessment manoeuvres were una eptable as
the demands aused ontrol surfa e rate limiting whi h is urrently not handled
in the RIDE ontroller. Time responses are already shown in Figs. 33.2 to 33.8,
Figs. 33.12 to 33.14 and Fig. 33.23.
- Test 31 failed as g and negative  limiting have not yet been implemented.
- Test 32 passed as shown in Figs. 33.24 and 33.25.

563
 Structural coupling, pitch rate to taileron () and to canard ()
5
10

0
10
Gain

5
10

10
10

15
10 2 1 0 1 2 3
10 10 10 10 10 10
Frequency (rad/sec)

Figure 33.24: Control law gain - pit h rate to ontrol surfa e

Structural coupling, normal acc. to taileron () and to canard ()

0
10

5
10
Gain

10
10

15
10

20
10 2 1 0 1 2 3
10 10 10 10 10 10
Frequency (rad/sec)

Figure 33.25: Control law gain - normal a eleration to ontrol surfa e

564
 - Test 41 did not produ e sensible results as the inputs aused the ontrol
surfa es to rate limit.
- Test 42 produ ed the following results.

Control a tivity due to turbulen e: 0.2807 RMS on symmetri taileron,

0.8031 RMS on dierential taileron,
0.5507 RMS on rudder.
Control a tivity due to noise was: 13.1636 RMS on symmetri taileron,
2.1729 RMS on dierential taileron,
13.1636 RMS on rudder.

33.7 Con lusions and Lessons Learned

A ontroller has been designed using Robust Inverse Dynami s Estimation
(RIDE) for the HIRM design hallenge. The RIDE method uses an estimate
of the inverse dynami s of the outputs to de ouple them. A Proportional plus
Integral (PI) ontroller is then added to assign the desired dynami s. The
proportional part of the PI ontroller provides stability and the integral part
ompensates for errors in the estimate of the inverse dynami s. A feedforward
element is in luded to shape the transient response of the losed-loop system.
The key advantages of the method are:
(i) the simpli ity of the ontrol law synthesis and theory, whi h only requires
a knowledge of elementary matrix algebra to fully understand,
(ii) an identiable ontrol law stru ture,
(iii) physi ally meaningful gain matri es and
(iv) the apability of handling motivator position and rate limiting in a
very simple manner. Motivator redundan y an be handled separately from
the ontroller design.
The overall simpli ity of both theory and design allows a rapid understand-
ing of the method, fast design and redesign, with modi ation being simple to
designers familiar with lassi al design methods.
The simpli ity of the design method and the resultant ontroller stru ture
does mean that only a subset of the design riteria an be in luded in the
ontroller synthesis. Only the response shape and bandwidth riteria map
onto the design parameters. Also, robustness guarantees are not in luded and
must be addressed separately using other tools.
An initial version of the ontrol law for the HIRM design hallenge was
synthesised very rapidly and with little modi ation, produ ed a eptable re-
sponses. A minimum amount of s heduling based on Ma h number proved
adequate and a minimum number of sensor signals were used. However before
denitive on lusions an be made about the suitability of RIDE, a more om-
prehensive ontroller needs to be designed and tested. Su h designs and tests
are already planned by DRA Bedford.

565
566
 Part IV

Con luding part

567
34. The Industrial View

C. Fielding and R. Lu kner

1 2

Abstra t. The results from the RCAM and HIRM design hal-
lenges whi h have been des ribed in the previous hapters were
evaluated in relation to industrial ight ontrol law design and de-
velopment pro esses. Experien ed evaluation teams were estab-
lished, led by Daimler-Benz Aerospa e Airbus for the ivil air raft
design results (RCAM) and by British Aerospa e Military Air raft
for the military air raft design results (HIRM). This hapter de-
s ribes the ommon evaluation pro ess used and the interpretation
of the results presented in the design teams' reports, in relation to
industrial experien e and pra ti es.

34.1 The Evaluation Pro ess

34.1.1 Fa tors ae ting the design of ontrol laws
In order for a team of designers to a hieve su essful ight ontrol laws design
and implementation, several fa tors will play a part, with the design te hnique
used being just one su h fa tor to be onsidered. In no parti ular order, some
of these fa tors are:

 The design team's experien e and knowledge of ight me hani s, ight

ontrol and ontrol theory.

 The tools and omputing fa ilities available to the design team.

 The time available to the design team and the level of support available.

 The adequa y and interpretation of the design requirements.

 The pro edures used for a hieving and ontrolling the design denition.

 The design team's experien e with the te hnique and the level of design
automation used.

 The visibility and larity of denition of the resulting design's fun tional
spe i ation.
1
British Aerospa e Military Air raft, Aerodynami s Department, Warton Aerodrome,
Preston PR4 1AX, UK
2
Daimler-Benz Aerospa e Airbus, Flight Guidan e and Control, Flight Me hani s Depart-
ment, P.O.Box 950109, D21111 Hamburg, Germany

569
 These are the main fa tors but there may well be others. Clearly it is always
going to be a di ult task to lter out the benets whi h an be attributed to
the design te hniques, from a set of results produ ed by dierent design teams
from dierent organisations. This was re ognised before evaluation ommen ed
and an evaluation pro edure established, aimed at produ ing the learest pos-
sible pi ture from the information available. This involved the development of
a questionnaire, to apture the signi ant information from evaluations of the
design entries (whi h were presented to the evaluators as a series of design re-
ports). The resulting questionnaire was ommon to RCAM and HIRM designs
and is summarised in the next se tion, so that readers an appre iate those
aspe ts whi h are important to industry and the type of questions that were
asked, and are likely to be asked in the future.

34.1.2 The Evaluation Questionnaire

Experien ed individuals from industry and resear h establishments were asked
to s reen and to evaluate the te hni al reports that were provided by the design
teams. Their evaluation was stru tured by a questionnaire whi h addressed the
following important topi areas:

 The eort to learn, to implement and to apply the method.

 The appli ability of the design method to ight ontrol laws design.

 The omplexity of the resulting ontroller, its implementation and erti-

 ation issues.

 The robustness and performan e of the designed ontroller.

Spe i ally, the following questions were asked, ea h supplied with a de-
tailed des ription to over the s ope and intention of the question, and a de-
s riptive rating (ex ept for the last three questions) s aled between 1 and 5:

1 (Question 1 was related to the quality of the design do ument)

2a What level of understanding of the methodology is needed to obtain sat-

isfa tory results ?

2b How do you rate the learning urve asso iated with using this method,
i.e. how easy is the method to grasp ?

2 How easy would it be for you to take over the design and arry out a
re-design, with the same method ?

2d Does the method support all ight ontrol law stru tures that you might
possibly want to design ?

2e Is it possible to translate all of your design requirements into the design

method syntax ?

570
 3a Do you onsider the ontroller stru ture presented to have good visibility
in terms of its fun tionality ?

3b How do you rate the omplexity of the design, in relation to the design
problem omplexity ?

3 How suitable is the design for implementation in an air raft's on-board

ight ontrol omputer ?

3d How suitable is the design for omplian e with your quali ation and
erti ation pro edures ?

4a Do you have any omments regarding the robustness of the design that
has been a hieved ?

4b Do you have any omments regarding the performan e of the design that
has been a hieved ?

4 Do you have any omments regarding the ontrol surfa e a tivity asso i-
ated with the design ?

In order to help to resolve any dieren es in judgement between the eval-

uators and designers, all designers arried out a self-evaluation of their own
design entry, by using the same questionnaire. Additionally, the questionnaire
identied the evaluators' and designers' levels of experien e in key areas of
responsibility within the ight ontrol laws development pro ess. This would
allow the ratings about dierent aspe ts of a design to be orrelated with the
experien e levels, and hopefully, to explain any areas of disagreement. Spe i-
ally, the following areas of expertise were assessed:

 Flight me hani s / Air raft stability

 Flight Control laws design

 Piloted simulation / handling qualities

 FCC implementation and testing

 Flight learan e / erti ation

 Flight test / post-ight analysis

In performing a design entry evaluation, the evaluators' judgements would

not only vary with their levels of experien e in these areas, but also on their
knowledge and experien e with the parti ular te hnique used to a hieve the
design. This aspe t was also addressed by the questionnaire.
Finally, it was planned to arry out an independent automati evaluation
of the various ontrollers by using software-based tools to give quantitative re-
sults. Unfortunately this information was not available in time for the evalua-
tion pro ess. The results of the automati evaluations will however, be available
in the updated design reports produ ed be ea h of the teams.

571
34.1.3 The Design Entries
In the following se tions, the design entries are often identied using a simple
notation:

Notation Organisation Design Method Appli ation Report/

Chapter

MS-11 DUT -Synthesis RCAM [25/22

EA-12 CERT-ONERA Eigenstru ture Assignment RCAM [55/19
CC-13 CUN Classi al Control RCAM [88/15
LY-14 LAAS-CNRS Lyapunov RCAM [49/20
FL-15 DUT Fuzzy Logi Control RCAM [207/26
MO-16 DLR Multi-Obje tive Parameter Synthesis RCAM [130/16
EA-18 LUT Eigenstru ture Assignment RCAM [77/17
MS-19 NLR -Synthesis RCAM [208/23
PC-20 UCAM Predi tive Control RCAM [119/25
HI-21 ULES H1 Mixed Sensitivity RCAM [242/21
EA-22 UNED Eigenstru ture Assignment RCAM [52/18
MF-25 DLR Model Following Control RCAM [67/24
CC-24 ALN Classi al Control HIRM [89
LQ-26 CIRA Linear Quadrati Control HIRM [9/28
MO-27 DLR Multi-Obje tive Parameter Synthesis HIRM
DI-28 DRA Robust Inverse Dynami s Estimation HIRM [73/33
MS-29 DUT -Synthesis HIRM [163/31
DI-30 ONERA Nonlinear Dynami Inversion/LQ HIRM [72/32
MS-31 SMA -Synthesis HIRM [225/30
HI-32 UCAM/CCL H1 Loop-Shaping HIRM [191/29
HI-33 ULES H1 Loop-Shaping HIRM [217

Where the two hara ter alphabeti identier indi ates the design method
and the two hara ter numeri al identier refers to the GARTEUR te hni-
al publi ation produ ed for the Robust Flight Control proje t; e.g. report
GARTEUR/TP-088-11 was produ ed for design entry MS-11 (any missing
numbers are asso iated with reports whi h were not overing design entries).
The expanded titles of the organisations an be found at page iii.

34.2 Civil Air raft Manufa turer's View 3

34.2.1 Introdu tion

The aeronauti al industry's motivation to look for new robust ight ontrol
4
design methods , is driven by e onomi al needs. The design pro ess has to
be ome more ee tive ( heaper, faster) and less risky (meet deadlines, be within
budget) while guaranteeing the same high safety level.

3
Prepared by Robert Lu kner on behalf of the RCAM Evaluation Team
4
Design method stands for the overall design pro edure.

572
 The introdu tion of new ontroller synthesis methods
5 into the industrial
development pro ess of ight ontrollers is done onservatively be ause the
urrent design methods represent a riti al apital for industry (trained people,
software tools, et .). New methods will be adopted, only if they are mature and
if benets have been demonstrated or if dynami ight ontrol problems that
annot be solved by existing methods, are driving te hnology. That is why, for
su h appli ations, it takes, typi ally 10 to 20 years for a new ontrol synthesis
method from its rst publi ation to its routine servi e in an industrial design
pro ess.
During the last twenty years, a faster appli ation was aggravated by the
fa t that design o es of ivil air raft manufa turers were o upied by the
introdu tion and exploitation of digital te hniques. Most of their resour es
have been absorbed to master the omplexity of digital ight ontrol systems
and their software. The design pro ess stands in the foreground and the sear h
for better ontrol synthesis methods was pushed into the ba kground as most
of the design problems ould be solved more or less satisfa tory by the broad
experien e and the extensive suites of lassi al design tools that have been
available. There were of ourse, improvements (e.g. appli ation of optimization
te hniques, transition from ontinuous to dis rete design methods) but this was,
more or less, an automation or a onversion of existing te hniques.
The tremendous development osts of digital ight ontrol systems and the
modi ation osts that are aused by hanges, even for only slightly dierent
air raft versions (e.g. dierent engine types, stret hed versions et .), brings
up new requirements whi h are not only physi ally, but also e onomi ally rea-
soned. In this ontext, the idea of robust ontrol is very attra tive. System
simpli ations or a redu tion of the number of dierent versions seem to be
feasible. If su h benets are foreseeable, traditional design pro esses should be
re onsidered and the introdu tion of new ontroller synthesis methods would
be justied.
The evaluation of new methods for ight ontrol appli ations is di ult
and ostly. Many dierent riteria have to be onsidered. Possible benets
have to be weighed against the additional eort, osts and risks. In the past,
de isions on the implementation of a new design method were often based on
the intuitive judgement of experien ed engineers. But as the omplexity of the
design pro ess has in reased and as the onsequen es of a de ision in today's
development pro esses an be dramati , a systemati , obje tive approa h is
needed.
The main obje tive of the RCAM design hallenge is the demonstration
of modern and robust ontrol synthesis te hniques to the European air raft
manufa turers, oering help for their de ision making.
Twelve design teams have demonstrated the appli ation of dierent design
methods to the RCAM design hallenge in the previous hapters. The RCAM
problem is a realisti , though simplied, ight ontrol example of industrial
relevan e. The teams have presented their individual ontroller stru ture, and

5
Synthesis method is the mathemati al method that is used to synthesize a ontroller
(LQG, H1 , Eigenstru ture Assignment, et .).

573
the results they have a hieved. Twelve of these methods are alled new as they
have emerged during the last three de ades and are only rarely used by the
ight ontrol o es of the aeronauti al industry. One team used a traditional
lassi al approa h, whi h an be taken to a t as a referen e.
The pro edure for the evaluation of the dierent ontrol design and synthesis
methods was set up, as previously des ribed. The relevant riteria for a design
method assessment were addressed systemati ally and evaluators from industry
and resear h establishments were asked for ratings and omments.
A pre ondition for the evaluation of dierent ontrol design and synthesis
methods was that all reports were written in a standardized format. This was
a hieved by des ribing the RCAM problem in a manual, dening the RCAM
model in MATLAB syntax with a standard nomen lature, and by pres ribing
a standard lay-out for the design reports.
The evaluation results, whi h are given and dis ussed below, show benets
and drawba ks of the dierent methods. They allow the formulation of well-
founded re ommendations, in order to ease the de ision making for industry's
design o es. But of ourse, the nal de ision to apply a new method has to
be made by the design o e itself.
It has to be noted that the RCAM hallenge serves only the above men-
tioned obje tives. It is not a ompetition between the design teams or their
organizations. There will be no winner and no loser.

34.2.2 RCAM Evaluation Team

The RCAM evaluation team onsisted of 20 evaluators, 12 from industry (Ale-
nia, AVRO, BAe-D, BAe-MA, DASA) and 8 from resear h establishments
(CERT, DLR, NLR). It was led by DASA. All evaluators are experien ed
aerospa e ight ontrol professionals. Colle tively, they represent a high level
of ompeten e in ight me hani s / air raft stability, ight ontrol laws design,
piloted simulation / handling qualities and they had experien e in FCC im-
plementation and testing, ight learan e / erti ation and ight test / post
ight analysis. As it was the suitability of a design method for the industrial
pro ess that was to be evaluated, an evaluator had not ne essarily to be fa-
miliar with the method he assessed. The questions of the questionnaire ould
be answered without this knowledge. In ase of doubt, questions ould be left
unanswered.
The evaluation was done on a voluntary basis. As a thorough evaluation
of a design entry takes one or two days, it was ne essary to divide the work
between multiple evaluators. Therefore, most evaluators evaluated one or two
reports and only one evaluator has evaluated all design reports. This has been
onsidered when the evaluation results were ompared.

34.2.3 Design Entries

The twelve teams that have presented their designs in the previous hapters
are from resear h establishments (CERT-ONERA, DLR Institute for Flight

574
Me hani s, DLR Institute for Roboti s and System Dynami s, NLR, LAAS)
and from universities (CUN, DUT Aerospa e Engineering, DUT Ele tri al En-
gineering, LUT, ULES, UNED, UCAM). They have used methods that dire tly
ope with robustness against dened un ertainties, i.e. the plant model used
for design ontains a nominal plant model plus expli it un ertainty modelling:

MS -synthesis MS-11, MS-19;

and they have used methods whi h, for the same a ura y of un ertainty mod-
elling, inherently provide some robustness measures (e.g. gain and phase mar-
gins, eigenvalue sensitivity) but where robustness has to be proven in an anal-
ysis, starting an iteration y le:

HI H1 HI-21,
EA eigenstru ture assignment EA-12, EA-18, EA-22,
CC lassi al ontrol design CC-13,
FL fuzzy logi ontrol FL-15,
MF model-following MF-25,
LY Lyapunov LY-14,
MO multi-obje tive parameter synthesis MO-16,
PC predi tive ontrol PC-20, PC-23.

The evaluation is based on the se ond revision of the design reports from
whi h the tutorial hapters and the design hapters of this book are derived.
Due to the tight time s hedule, an evaluation of LY-14 and PC-20 ould not
be in luded in this hapter.

Designer's Ba kground
The prin ipal designer's level of experien e in dierent areas of ight ontrol
system development was determined by the self-assessment. The self ratings
state his level of experien e at the beginning of the design and do not des ribe
the ompeten e of the resear h establishment or university and they do not
in lude the support that a designer had available within his organization.
All design teams knew the synthesis method they have applied quite well.
They rated their knowledge and experien e as medium (CC-13, FL-15, EA-18,
MS-19, HI-21) and high (MS-11, EA-12, MO-16). Only the designer of MF-25
used the method for the rst time.
Three design teams (FL-15, MO-16, HI-21) rated their level of expertise
in ight dynami s as low, i.e. very little to basi knowledge. This expertise
is important for the sele tion of the ontroller stru ture, for weighting design
requirements (whi h are often in ontradi tion) and for understanding and
interpretation of the dynami behaviour of the ontroller.
It is not surprising that the knowledge on theoreti al subje ts su h as ight
me hani s, ight ontrol laws design and simulation is mu h higher than on
pra ti al subje ts as implementation, testing, and erti ation. This an be
even an advantage, as too mu h knowledge of possible implementation and
erti ation problems involves the danger of being too onservative and too
relu tant in applying new methods.

575
Control Strategy and Ar hite ture
A lot of the ontroller performan e is predetermined when the strategy and the
ar hite ture is dened. Tables 34.1 and 34.2 give an overview of the ontroller
stru tures hosen by the design teams.
Basi ally, traditional ight ontrol ar hite tures have been hosen, i.e.:

 De oupling of the ontrollers for longitudinal and lateral motion, some-

times with oupling terms (e.g. for turns). An ex eption is design HI-21
whi h ombines longitudinal and lateral inner ontrol loop into one dy-
nami blo k. This is a quite un onventional stru ture but it promotes
very interesting alternatives.

 Hierar hi al stru tures with an inner loop for stabilization and attitude
ontrol and an outer loop for guidan e. Ex eptions are MS-11, MO-
16, and EA-18 whi h ombine inner and outer loops in the longitudinal
ontroller.

It is well known, as many evaluators have emphasised, that a suitable hoi e

of the ontroller ar hite ture, based on ight physi s and mission demands is
mu h more relevant than the applied synthesis method. To establish su h an
ar hite ture, a lot of experien e and ight dynami s knowledge is ne essary.
For designers with very little or only basi knowledge in ight dynami s (FL-
15, MO-16, HI-21), the sele tion of signals and ar hite ture was a very steep
hurdle that they had to over ome. They managed it with some iteration and
with the help of ight ontrol textbooks.
The RCAM manual oers diverse referen e signals and measurement signals
as ontroller inputs. The designers sele ted the input signals for their ontroller,
individually oming to dierent solutions. Some ommonly used measurements,
that are available from the inertial referen e systems, are not oered, e.g.  and
. Estimation of missing measurement signals, as done by some designers, is
onsidered as a part of the ontroller. This aspe t ould have been assigned to
the RCAM air raft model. Thrust, elevator, aileron, and rudder were available
for air raft ontrol and all of them were used by all designers.
Most of the designs suer from the inability to formulate referen e values,
whi h are onsistent with ight physi s and mission requirements. The 9 refer-
en e signals ( x , y , z , uV , vV , wV , V ,
y , _ ) dene a non-yable ight
path for the RCAM mission: entry and exit of turn requires innite roll rates
and the transition from level ight to des ent requires an innite verti al a -
eleration. A few teams used models (MS-19 and MF-25) or prelters (MS-11,
MO-16) to generate a yable path, others introdu ed mode swit hes (CC-13,
EA-18) and limiters (CC-13, FL-15, MS-19), feedforward (EA-18).
While the denition of the overall (high level) stru ture was nearly inde-
pendent of the sele ted synthesis method (ex ept MF-25), only the low-level

6
inner loop order + outer loop order = total order
7
The 26th order inner loop ontrol blo k is used for longitudinal and lateral ontrol

576
 long. outer long. outer long. inner long. inner remarks long.
loop loop loop loop ontroller
6
ommand feedba k ommand feedba k order
MS-11 h, VA outer and inner loop are ombined; prelter: 24
h, h_ ;
feedba k: h, h_ , VA , V_ A , q ,  VA , V_ A
MS-19 z , wV z , wV wV , VA , q wV , VA , q model, rate 4+16=20
limiter
HI-21 z z wV , VA wV , VA , q mode swit h 5+(26)
7

FL-15 h, h_ , h, h_ , , DT H q,  limiter 2
VA V , nx R
EA-12 z z R wV , V , wV , R wV , (zg , wV;g ) ! 2+4=6
(RwV -wV; ), V , V , R wV;
(V -V ), nz , q (RwV -wV; ),
q oord R (V -VR )
EA-18 z , wV , VA z , wV , VA , no inner loop VA , z 2
q , nz R mode swit h
EA-22 z , wV z wV , VA wV , R wV , 3
VA , VA ,
nx , nz , q
CC-13 z , wV , VA z , wV , VA  , VA , q,  limiter, mode 4+1=5
q oord R sele tion
MF-25 z , wV , VA z , wV , VA uB , w B , q, q model 8+1=9
u_ B , w_ B ,
q oord
MO-16 z , VA z , VA , q , no inner loop mode swit h 4
nx , nz , wV
Table 34.1: Overview of Longitudinal Controller Stru tures

stru ture was inuen ed by the sele ted synthesis method.

Evaluation of Design Method and Controller Stru ture

The evaluators' ratings and the omments on the questions on erning the
eort ne essary for learning and appli ation of the design method (question 2a
to 2e), on erning the omplexity of the ontrol solution (question 3a to 3d)
and the general omments, have been used for the following dis ussion of the
design method and ontroller stru ture. More details are given in the RCAM
evaluation report [103.

H1 and -Synthesis
The three approa hes to robust design that are based on H1 design (HI-21)
and -analysis (MS-11, MS-19) are grouped as they lead to similar evaluation
8
inner loop order + outer loop order = total order
9
The 26th order inner loop ontrol blo k is used for longitudinal and lateral ontrol

577
 lat. outer lat. outer lat. inner lat. inner remarks lat.
loop loop loop loop ontroller
8
ommand feedba k ommand feedba k order
MS-11
y
y ,  ,  ,  , p, r prelter: 8+24=32
MS-19 uV , vV , uV , vV ,  , p, ,  , p, model 3+15=18
_ ,
y r , vV r , vV
HI-21 Vx , Vy ,
y ,   ,  , p, r model, 2+3+(26)
9

y ,  mode swit h
FL-15
y ,  ,  R  , p, r Rlimiter
R 4
EA-12
y ,  ,   , R ( - ), ,  , p, r ,  3
, (- )
EA-18
y , _ ,  , p, no inner loop
R R
, vB 3
r, y ,  R R
EA-22
y ,  ,  ,  , p, ,  2
r, 
CC-13
y ,   , p (roll), limiter, 2+3=5
,  , r (yaw) mode
sele tion

MF-25
y , _ R,  , p, r, , R, p, model 14+1=15

 , p_, r_ r, 
MO-16
y , _  ,  , p, r prelter, 3+2=5
mode swit h

Table 34.2: Overview of Lateral Controller Stru tures

results.
HI and MS rely strongly on modern mathemati al ontrol theory. As a
normal ight ontrol engineer in industry is not familiar with HI and MS,
it will take him time and eort to gain the understanding and the experien e
whi h is ne essary for a professional appli ation (6 to 12 months). Knowledge in
lassi al frequen y-domain te hniques, that an be presumed, will ease learning.
Another positive aspe t is that ommer ial software is available. All three
design teams used ommer ial produ ts (MATLAB), without undue di ulty.
Experien e is ne essary in how to translate design requirements into fre-
quen y dependent weighting fun tions. The designers used model-following
te hniques to in orporate most of the design requirements into ideal models.
This approa h left only a few requirements for translation into HI or MS syntax
(i.e. weighting fun tions). A model-following approa h is not inherent or par-
ti ular to HI and MS; it has been ommonly and su essfully used by industry,
e.g. for autopilot designs.
HI and MS in lude dened plant un ertainties by an un ertainty model
 guaranteeing robust stability and robust performan e . Note that the term
performan e in the method's sense, does not over all the RCAM performan e
requirements, nor typi al industrial performan e requirements given in a ight
ontrol system spe i ation. The un ertainty model is a relevant improvement

578
with respe t to a systemati design, but un ertain parameter ranges are to
be spe ied expli itly, and therefore they have to be known in advan e. An
ex eedan e of the dened and modelled un ertainty range, might deteriorate
ontroller performan e drasti ally. So, the un ertainty modelling has to be done
with are and physi al understanding. Only limiting values that orrespond
with physi al boundaries or that must not be ex eeded in air raft operation
(e.g. maximum weight) an be dened easily.
Also some aution on erning the delity of the un ertainty modelling is
advisable. The un ertainty model represents a set of linear models and annot
over spe i nonlinearities of the plant. Although this does not present a
major drawba k for many ight ontrol design tasks, it does require additional
testing, as there is no impli it proof for stability of the nonlinear system.
The presented HI and MS design pro esses onsist of multiple steps, in lud-
ing analysis steps with linear and nonlinear models. As not all requirements
an be formulated in a straightforward manner, iterative tuning is ne essary.
The tuning was done manually without support of an optimizer.
The presented ontrollers are of high order. The ontroller dynami s of
MS-11 turned out to be ill- onditioned during the design pro ess (fast poles,
unstable poles), but an improvement was possible with some extra eort. Un-
stable ontrollers are una eptable in general, and marginally damped and high
frequen y poles ause problems during transformation of the ontinuous design
into a dis rete ontrol algorithm. Order redu tion and residualisation has been
used as a remedy. The evaluators ould not assess the ontrollers' dynami s,
as the eigenvalues of the ontrollers are not given in the design reports.
Only a part of the ontroller, but a big and the most important blo k, has
been designed by HI and MS ontrol te hniques. It seems to be impossible
to gain physi al insight into the dynami s of this blo k  espe ially if order
redu tion te hniques are applied. This blo k has to be looked at as a bla k
box. Its missing visibility and its high omplexity presents a risk for the
industrial learan e and erti ation pro ess. Solutions need to be found
10 .
Another risk is the possibility of an abrupt ontroller performan e deterio-
ration, if the assumed un ertainty range is ex eeded. A ight ontroller for a
modern transport air raft will be embedded into 1000 to 2000 air raft whi h
will operate 100 million ight hours over 50 years. History has shown that
they will fa e many unpredi table events and situations. The ight ontrol
system should behave a eptably, as lassi al low gain designs typi ally do.
If ontroller's deterioration an be predi ted or dete ted  preferably without
additional sensors, a re onguration an be initiated.
Results of the HI and MS design entries are amongst the best. One fa tor
is the higher potential of the high-order ontroller, another fa tor is that the
methods are well tailored to handle the RCAM model un ertainties. The fa t
that ontrollers of signi antly lower order have rea hed similar results, leads
to the assumption that drasti order redu tion is possible.

10
From a mathemati al point of view the omplexity of this blo k is low but the resulting
algorithm for the embedded software and the extensive testing that is ne essary leads to the
erti ation issues.

579
 The on lusion is: HI and MS are more powerful te hniques than the las-
si al ones. They oer an interesting potential  whi h has to be paid for.
Weighing the pros and ons of HI and MS design for an industrial ight ontrol
task has to be done ase by ase, e.g. additional eort might be justied if
sensors an be saved or the number of operating modes an be redu ed.

Fuzzy Logi Control

Evaluation results are based on entry FL-15.

FL belongs to the group of intelligent ontrol te hniques and was introdu ed
in 1965 by Zadeh. The power of FL is to design ontrollers in an intuitive
way. Design is done by linguisti formulations and reasoning. A eptable but
not ne essarily optimally tuned solutions are the out ome. A designer has to
understand the general idea of FL whi h is easy to grasp and he needs exer ise
in appli ation of FL. Good knowledge of ight physi s and piloting te hniques
is ne essary but an be taken as granted for an average ight ontrol engineer.
Therefore, the eort for learning has been rated moderate to easy. As the
method yields a nonlinear gain ontroller, theoreti al knowledge in nonlinear
dynami systems ould be useful.
A variety of ommer ial software tools are available. In the presented design,
the author used a MATLAB toolbox that was developed at Delft University.
FL seems to have its potential in improving those parts of a ight ontroller,
where a ontrol strategy an be opied from piloting te hniques. So it is quite
logi al, that a hybrid design te hnique was hosen: the outer loops were de-
signed by means of FL, while for the inner loops lassi al design te hniques
have been applied.
It is very easy to formulate linguisti riteria in FL syntax, but it is not lear
how to expli itly introdu e, quantitative frequen y-domain riteria. Further-
more, the FL method does not address robustness dire tly. Robustness an be
addressed indire tly: if the ontroller uses low gains and if the behaviour of a
human ontroller is mimi ked, as a human ontroller is normally adaptive and
robust. But it has to be proven by analysis that robustness is a hieved, bring-
ing the FL design methods to the typi al multi-step, iterative design pro ess
with trial and error.
A signi ant advantage of FL is that linearization of the ight me hani s
equations is not required for the design, but linearization was ne essary to nd
gains for the lassi al design of the inner loop ontroller.
The fuzzy part of the ontroller is a stati , nonlinear, deterministi fun tion
that an be realized as a lookup table. The ontroller stru ture is visible and
the order is very low. No implementation and erti ation problems are to be
expe ted.
For the implementation of FL ontroller algorithms into a ight ontrol
system by using a nonlinear look-up table, erti ation should be no problem. If
fuzzi ation and defuzzi ation software or fuzzy opro essors are used instead,
questions regarding erti ation need to be answered.
The FL design te hnique aims at dynami systems whi h are di ult to
model and that are hara terized by un ertainty and impre ision. This does

580
not apply to ivil transport air raft dynami s. They are a urately modelled
(from the FL point of view) with mu h eort and the models whi h are used
for design and pilot training purposes are improved after ight testing by pa-
rameter identi ation te hniques. The situations where un ertainties an o ur
happen during ight testing and in rare failure ases.
A bene ial future appli ation of FL ould be in the area of ight guidan e,
where piloting te hniques an be opied and where pre ision requirements are
not too high, e.g. relaxed altitude ontrol during ruise in order to take are
of the engines.

Eigenstru ture Assignment

Evaluation results are based on entries EA-12 (Modal Multi-Model Control

Approa h), EA-18 and EA-22 (Eigenstru ture Assignment).
EA is an extension of the well-known pole-pla ement te hniques. EA al-
lows the designer to assign the losed-loop eigenvalues (poles) and additionally,
to assign the eigenve tors or parts of them, within ertain limits. Desirable
pole lo ations are spe ied in the military ying quality spe i ations (MIL-F-
8785C). The military spe i ations are often used as guidelines for ivil trans-
port air raft, though their fullment is not ne essary for erti ation. By the
assignment of eigenve tors, the zeros of the transfer fun tions an be inuen ed
and oupling and de oupling of states and modes an be addressed dire tly.
Though EA an be learned with a eptable eort, very good knowledge of the
pro ess and its eigenstru ture is needed.
EA is available in ommer ial standard software pa kages. All three de-
sign teams used MATLAB. EA-12 used its CERT in-house developed MODAL
CONTROL TOOLBOX, EA-18 and EA-22 used their own MATLAB M-les.
EA methods an design state feedba k or output feedba k, with onstant
gains or with dynami feedba k. This allows a great variety of possible on-
troller stru tures. EA-18 and EA-22 designed stati feedba k ontroller. EA-12
used stati gains for the lateral ontroller and a fourth order dynami ontroller
for the longitudinal axis.
EA is strongly linked to linear systems. A potential is seen for exible
air raft where the aeroelasti behaviour is typi ally des ribed in terms of modes.
De oupling of aeroelasti modes or minimum damping values are important
requirements. The design task is aggravated by marginal ontrollability of
ertain modes by the standard aerodynami ontrol surfa es.
It seems to be impossible to dire tly formulate riteria su h as a tuator
a tivity or passenger omfort as onstraints on variables; it is however, possible
to do this indire tly, e.g. by hoosing losed-loop eigenvalues to ensure roll-o
in order to inuen e ontrol surfa e a tivity. Work-arounds were used, ending
with an iterative multi-step sear h pro edure. Guidelines for the denition of
referen e eigenve tors ould be worked out, possibly by analyzing other good
designs. The few hints given in the reports on how the eigenstru ture should be
modied in order to a hieve those problemati al requirements are not su ient.
In order to address all design riteria, it may be ne essary to supplement
EA with other te hniques, as demonstrated by the design teams.

581
 The time ve tor method of Doets h [54, where the solutions of dierential
equations for for e and moment equilibrium are plotted for ea h one of the
eigenve tors, ould be a possibility to ombine mathemati s and physi s in
order to gain better insight  but some arduous work has to be expe ted.
The three teams dened low-order ontrollers that have a rather onven-
tional stru ture with good visibility, i.e. implementation and erti ation an
be done by standard pro edures. Robustness was a hieved by using low feed-
ba k gains (whi h gave adequate performan e) whi h were then he ked in a
multi-model analysis.

Classi al Approa h

Evaluation results are based on entry CC-13.

CC is well known in the aeronauti al industry and has been applied sin e
the beginning of automati ight ontrol. Chara terized by su essive loop
losure, the lassi al design is guided by understanding of the physi s of ight
and a good deal of intuition and experien e that assists in sele ting ontroller
stru ture. Physi al insight is ru ial and inherent to the method. CC is not a
robust ontrol design method. Nevertheless, the presented design shows that
robustness in the sense of the RCAM design hallenge an be a hieved with
lassi al design methods.
A wide variety of ommer ial software tools are available. In the presented
design, the author used MATLAB.
This design entry is very useful as a ben hmark for the other entries using
modern methods. It demonstrates that an a eptable solution is possible by
some engineering eort. The presented ontroller stru ture is visible and the
order is low. No implementation and erti ation problems are to be expe ted.
If CC leads to satisfa tory results, one might ask: "why look for modern
robust methods?". There are at least two reasons:

 The eort to apply CC (SISO te hniques) in reases progressively with the

omplexity of the plant model. CC omes to its limits when ontrollers
for MIMO systems with high internal oupling are to be designed. That
is why MIMO ontroller synthesis te hniques (e.g. LQG) are applied
in industry. For future demanding design tasks, su h as an 800-seater
air raft or a new supersoni ommer ial transport, powerful and robust
te hniques are required.

 The di ulty to implement CC into a systemati and highly automated

industrial design pro ess, as the designer and his experien e is an integral
part of the method. This is underlined by the statement of one evaluator:
a CC re-design ould be a nightmare, if ... the rationale behind the design
is not available. This situation presents an in al ulable and potential risk
for a modern industrial design pro ess.

Model-Following Approa h

Evaluation results are based on entry MF-25.

582
 The MF approa h onsists of three fundamental blo ks that hara terize
the overall ontroller stru ture:

 A ommand blo k (CB) whi h denes the required dynami behaviour

(model) and whi h generates ideal state variables and their time deriva-
tives as ommand signals.

 A feedforward ontrol blo k (FFC) whi h ontains an inverse air raft

model. The FFC blo k omputes ontrol surfa e dee tions from the
CB's ommand signals. With an a urate inverse model and no distur-
ban es, the air raft (plant) would exa tly follow the ommanded signals
and all requirements that are inherently dened into these signals would
be fullled dire tly.

 A feedba k ontrol blo k (FBC) whi h ompensates for dieren es be-

tween a tual and ommanded signals aused by disturban es (wind, sen-
sor noise, et ), and un onsidered plant un ertainties.

The CB is asso iated with the outer loop, the FFC and FBC blo ks belong
to the inner loop.
To use MF for in-ight simulation seems to be very reasonable and has
been done su essfully for many years. Appli ations of the MF ontrol system
design as presented in MF-25 are not known for automati or manual ight
ontrol, though it is urrent pra ti e to use model-following for partial tasks of
autopilots (e.g. altitude a quire mode or are mode).
The lear task sharing and the underlying philosophy makes it easy to grasp
the general MF design on ept for any ight ontrol engineer with a ba kground
in ight me hani s. The eort for learning MF was rated moderate. The CB
synthesis an be straightforward, but experien e in the denition of a dynami
system that fulls design requirements is needed. Re-use of a suitable CB is
possible, i.e. it an be used for any air raft that has to full a set of existing
requirements. In the same way, previous designs with desired performan e an
be used for the CB. The authors took advantage of this feature: they based
their CB on the ontrolled RCAM model from CC-13
11  whi h fulls the
design requirements.
The prin iple, to put all knowledge on system dynami s into the feedforward
path (FFC), is taken into a ount by inversion of the nominal linear plant
model. The matrix algebra involved in this al ulation is only at a basi level
and not di ult. The authors used simpli ations that rely on the assumption
of ideal model-following. The question, what happens if this assumption is
not valid, has not been answered theoreti ally.

11
MF-25 took advantage from its late entry time, where rst results of other teams were
available. As the design of the CB an be seen as an isolated task, this approa h helped in
meeting the tight time s hedule. This is paid for by a predened (higher than ne essary)
omplexity of the CB and the a hievable results whi h has to be the same as for CC-13. The
results for CC-13 and MF-25 omplied with requirements.

583
 A simple time-domain optimization te hnique is used to design the low-order
feedba k ontroller. The authors point out that more sophisti ated synthesis
methods an be used, if required.
The author used MATLAB for his design, omplemented by an optimization
subroutine.
The MF approa h splits the design requirements into 2 ategories and adds
a third:

1. Requirements that an be designed into the CB. These requirements on-

ern the desired ommand behaviour and are fullled by means of the
CB and FFC blo k.

2. Requirements that have to be fullled by means of feedba k, e.g. stability,

ontrol a tivity, passenger omfort and disturban e reje tion riteria.

3. The additional MF requirement that the FFC should provide deviations

to be as small as possible between ommanded and a tual response.

Robustness in the sense of robust ontrol theory is related to feedba k.

Therefore the question of robustness arises only for the requirements under 2.
Many of the RCAM design requirements (e.g. ight path tra king) belong
to ategory 1. They have been dened dire tly into the CB and the expe ted
robust fullment has been a hieved and demonstrated.
The question of robustness be omes relevant for FBC design. The synthesis
method used does not onsider robustness expli itly. An additional time delay
has been introdu ed, probably to keep gains and frequen ies low. Robustness is
demonstrated by validation in worst- ase simuations. The ontroller appears to
be generally robust. In some o-design ases, the damping of higher frequen y
modes tends to de rease.
Complexity of the MF-25 design, in terms of order, was rated moderate
though it is higher than ne essary, be ause the ontrolled RCAM model from
CC-13 has been used in the CB. A simpler CB stru ture seems to be feasible.
Feedforward omplexity is determined by the omplexity of real air raft dy-
nami s. The inversion of the nominal linear air raft model produ ed a stati
FFC blo k of low omplexity. The longitudinal and lateral FBC blo ks ea h
have a low order of unity.
The high level of visibility and lassi al nature of this design should lead to
straightforward erti ation. No problems are expe ted.
The a hieved ight path tra king results are similar to CC-13. As CC-13
was sele ted as the model in the CB this is not surprising.
The importan e of the exa t knowledge of the air raft dynami s for the
FFC design is emphasised. Therefore, it would be interesting to see what will
happen, if the FFC blo k does not exa tly mat h the real air raft dynami s
(failure ases, unmodelled aeroelasti modes, operation at o-design points,
et .). Are more sophisti ated (nonlinear) inverse models needed? Then the
omplexity of the design an in rease signi antly.
MF proved to ope easily with the design hallenge. The lear strategy and
philosophy makes it very attra tive. Some questions on erning the robustness

584
 apabilities and the limitations of MF in more demanding design tasks have to
be answered.

Multi-Obje tive Parameter Synthesis

Evaluation results are based on entry MO-16.

The MO approa h onsists of a omputer-aided sear h pro edure, where the
designer determines his best ompromise between a number of spe ied rite-
ria. An ee tive division of work is organized, where the design engineer takes
are of all work that requires human intelligen e (i.e. to make de isions and to
be reative, e.g. in sele ting the ontroller stru ture, in hoosing synthesis te h-
niques, omposing the analysis hain, and de iding whi h ompromise is best),
and the omputer takes are of routine work (parameter optimization a ording
to given ost fun tions, data management, et .). The approa h is independent
of the ontrol synthesis te hnique or the sele ted ontroller stru ture. MO an
take any requirement into a ount that an be expressed mathemati ally by a
ost fun tion.
The learning eort is di ult to rate. Greater knowledge of the supporting
software environment ompared to the other design entries seems to be ne es-
sary, espe ially to implement ontrol design methods, to introdu e ost fun -
tions and to organize the automated design pro ess. A division of labour an
be organized between a omputer software spe ialist, who sets up the software
environment and adapts it to the design task, and the ight ontrol spe ialist
who is doing the ontroller design. This will relieve the workload of the ontrol
design engineer signi antly, making best use of his spe ialized apabilities.
The author used the DLR-proprietary design environment ANDECS with a
database fa ility. Commer ial ontrol software su h as MATLAB and MAT RIXX
an be oupled via pro ess ommuni ation. A good CAE software environment
in ombination with a powerful omputer (PC or workstation) is absolutely es-
sential for this design approa h.
The authors used lassi al and LQR design as low-level synthesis te hniques.
The presented ontroller stru ture is visible and its order is low. No implemen-
tation and erti ation problems are to be expe ted. Mode swit hing e.g. be-
tween the altitude mode and the limb/des ent mode has been used to over the
varying operating onditions. The ee ts of dierent dynami al designs an be
seen in the time histories. Robustness is a hieved by a worst- ase multi-model
approa h.
The MO approa h appears to be very attra tive for industry, as it ree ts
and automates typi al design pro edures. The freedom to in orporate any
riterion and any low-level synthesis te hnique is a big advantage and allows
design of all stru tures a designer wants to design. This is ree ted in the high
ratings for questions 2d and 2e. Depending on the omplexity of the models
and the ontrollers, omputing time for parameter optimization may be ome a
relevant fa tor.
The graphi al user interfa e that supports the designer has a signi ant
inuen e on the e ien y of the design pro ess. A good presentation of the
relevant requirements is essential for his de ision task.

585
 Flight ontrol systems are very omplex, and multiple requirements have
to be visible for weighting them simultaneously. The dimension and resolution
of a omputer s reen are limited. Therefore, the set-up of the graphi al user
interfa e has to be organized e iently, whi h will require some additional
eort.

Summing it up, MO has a great potential to help stru turing and automat-
ing the industrial pro ess for ight ontrol design.

34.2.4 Dis ussion of the Evaluation Results

Design Methods
The term design method has not always been used by the design teams in
the same sense: sometimes it was limited to the ontrol synthesis method
(MS, H1 , EA, LQR, et .) and sometimes it was used for the overall design
method that omprises multiple design steps. Due to this linguisti ina ura y,
the presented methods are named after the underlying mathemati al method
(design entries MS-11, EA-12, FL-15, EA-18, MS-19, HI-21 and EA-22). Only
three are labelled after hara terizing features (design entries CC-13, MO-16
and MF-25). The design teams used three dierent approa hes to assemble the
design method:

1. Control theory approa h (MS-11, EA-12, FL-15, EA-18, MS-19, HI-21,

EA-22):

In the entre stands the ontrol synthesis method  preferably a powerful

one. The ontroller stru ture has to omply with the onstraints of the
method. Flight dynami s and mission requirements have to be translated
into the method syntax  often a di ult task.

2. Flight physi al approa h (CC-13, MF-25):

Flight dynami s, mission requirements and physi al reasoning are en-

tral. Controller stru tures are build up in blo ks that orrespond to
physi al fun tions or pilot tasks. Control te hniques are hosen that an
be easily adapted to the design task. Optimization te hniques are used,
half-automated tuning and trial and error pro edures, too.

3. Pro ess oriented approa h (MO-16):

This approa h fo uses on the design engineer and the design pro ess, mak-
ing use of the automation potential that is provided by today's powerful
omputers (PC and workstations) and software. An ee tive division of
work is organized, where the design engineer takes are of all work that
requires human intelligen e (i.e. to make de isions and to be reative),
and the omputer takes are of routine parameter sear h work.

586
Air raft Model for Flight Control Design
The RCAM Model is a nonlinear model dened by nonlinear dierential equa-
tions. All designers had to linearize it, be ause their design te hnique is appli-
able to linear models only; the FL design method is the only ex eption. The
nonlinear model was used for analysis and tuning.
The MS design method requires the expli it modelling of the un ertainties.
This additional eort is honoured with guaranteed robustness on erning sta-
bility, within the bounds of the design envelope and the assumed parameter
un ertainties.
Methods like MS and HI, that design high-order ontrollers whi h may
a hieve better performan e, often design on the point, i.e. performan e dete-
riorates in o-design ases (i.e. failure modes, negle ted dynami ee ts et ).
This requires high-delity modelling of the air raft dynami s and exa t knowl-
edge of un ertainties. It is quite ommon that models of the air raft dynami s
have to be orre ted during ight testing, whi h ould lead to re-design and
modi ation of the ight ontroller in a late proje t phase. Subsequent ad-
ditional testing implies a potential risk in terms of budget and time for an
industrial programme. Certainly, this fa t applies to all methods, but the risk
has to be rated higher for the high-performan e MS and HI ontroller designs
than for less apable, but often well-behaved, lassi al designs.

Evaluation Questionnaire (Overall View of Entries)

The following dis ussion is based on the mean values and the distribution of
the ratings. In general, the results of the independent evaluators show good
agreement, varying opinions ould always be explained by the dierent inter-
pretation of a question. Only the key points are addressed below.

Eort Ne essary for Appli ation

Level of understanding (Question 2a). The eort to gain the ne essary level
of understanding was estimated to be medium for EA and FL. Higher eort
is required for HI and MS to understand the mathemati al ba kground and to
gain experien e in the method-dependent denition of requirements. MS needs
additional knowledge to model (stru tured) un ertainties by LFT. The ne es-
sary level of understanding for MO depends strongly on the sele ted synthesis
te hnique (e.g. LQR for longitudinal ontroller and CC for lateral ontroller).
The additional eort to ongure the software tool for MO is dependent on its
user-friendliness and ould not be rated. CC requires detailed knowledge, but
it an be presumed that mu h experien e from previous designs is available in
industry.

Learning urve (Question 2b). Similar results were obtained for the learning
urve. FL and EA an be understood with moderate and a eptable eort,
while MS and HI are more di ult to grasp.

Re-design (Question 2 ). A re-design with a method you are not familiar with is
always likely to be problemati al. It is di ult to estimate how many problems

587
you will en ounter without starting the work. This holds for lassi al and new
methods, and seems to be the reason why all design methods are rated, more
or less, in a small band around normal eort.

Flight ontrol law stru tures (Question 2d). The denition of the ontroller
stru ture and the denition of its internal fun tional blo ks (ight path gen-
erator, inner and outer loops, de oupling of longitudinal and lateral axis) is
essentially based on the interpretation of the requirements, physi al insight
and experien e in ight ontrol design, rather than on the method. The hoi e
of the synthesis method often poses unwanted onstraints on the sele tion of the
ontroller stru ture. Therefore, EA, FL, MS, and HI are used only for ertain
parts of the overall stru ture. Controller parts where these synthesis methods
were not appli able for design, have been synthesized by other methods. MO is
appli able for every stru ture that an be des ribed mathemati ally; this is why
it re eived the highest ratings. Entry HI-21 proposes an interesting ombina-
tion of longitudinal and lateral inner loops, leaving oupling and de oupling to
the synthesis method. It would be interesting to ompare the high-order on-
troller to a onventional approa h where oupling and de oupling are dened
by stru tural blo ks (e.g. turn ompensation).

Design requirements (Question 2e). All methods have shown their power when
design obje tives and method t; when otherwise, they have degraded to simple
trial and error. In MO every requirement an be used that an be expressed
mathemati ally by a ost fun tion. That is why it re eived the highest ratings.

Complexity of the Control Solution

Visibility (Question 3a). The visibility of all approa hes is a eptable or better.
The MS methods have been rated signi antly lower, at the limit of being
unsatisfa tory, be ause of their high-order dynami blo k for whi h physi al
reasoning seems to be impossible.

Complexity (Question 3b). Parts of the stru ture that make it omplex (non-
linearities, swit hes, lters, et .) are method independent. Other parts, su h
as high-order dynami blo ks are learly related to MS and HI. That is why
the omplexity ratings for MS and HI are at the limit of being unsatisfa tory.
Performan e omparisons showed, that omparable results have been a hieved
with ontrollers of mu h lower order.

Implementation (Question 3 ). The experien e of the design teams in imple-

mentation of ight ontrollers into a ight ontrol omputer was very low. Not
surprisingly, implementation issues have not been dis ussed in depth. There-
fore, evaluators' omments are more general.
All designs are ontinuous-time ontrollers. Typi al sample times for ivil
ight ontrol systems are 25 ms for the inner loops and 50 to 150 ms for the
outer loops. This poses restri tions on ontroller bandwidth. Furthermore,
dis retization of high-order ontrollers (HI and MS) an ause di ulties: high
omputing load or numeri al integrity of the dis rete algorithm, espe ially when
the ontroller has lowly-damped modes. Disturban es of measurement signals,

588
su h as bias or noise, were not onsidered. Smoothing lters were not used.
1
The use of ideal integrators ( s ) will ause una eptable drifts in real world
systems. Most of these implementation issues an be solved, but they may lead
to additional iteration steps.

Certi ation (Question 3d). Certi ation issues have been addressed in general
terms, for the same reason as for the implementation issues. Problems are
expe ted with erti ation of the high-order dynami MS and HI ontrollers.
As long as FL ontrollers are realized by nonlinear gain maps instead of FL
algorithms, problems are not expe ted. If a real performan e benet an be
demonstrated it is probable that methods for quali ation and erti ation
would follow.

Controller Performan e
Controller performan e ould be evaluated only qualitatively as the quantita-
tive results of the assessment software were not available during the evaluation
phase. Therefore, a more general dis ussion follows.

Fullment of Requirements

All entries fullled, more or less, the minimum requirements, as judged from
the plots in the design reports. If fullment is not omplete, it ould have been
a hieved probably with a little bit more tuning eort or slight modi ations of
the ontroller stru ture.
The design teams used dierent approa hes to generate the ight path om-
mands, whi h are used as ontroller inputs. This inevitably leads to method-
independent performan e dieren es and it be omes ne essary to lter out
this ee t. This variation is aused by the missing denition of a ight path
generator in the RCAM des ription.
The hoi e of stru ture and order was left to the design teams' dis retion.
Obviously, the more omplex a ontroller stru ture is (dynami s, lters, logi ,
nonlinearities), the better are the a hievable results. Controller order lies be-
tween 5 (EA-12) and 56 (MS-11). Comparably good results  as far as ould be
judged from the evaluated results  have been a hieved with 9th and 36th order
ontrollers. It is fair to assume therefore, that the high-order HI/MS ontrollers
an be onsiderably redu ed in order, without ompromising performan e. If
onrmed, order redu tion would be mandatory before implementation into
the ight ontrol omputer.

Robustness

There is no universal ontroller design te hnique that an over all design re-
quirements, in luding the robustness requirements, in one step. Therefore, all
teams built up design methods around their basi synthesis method. Iterative,
multi-step design pro edures with synthesis and analysis steps are the out ome.
Controller tuning is performed by means of sear h methods, i.e. optimization
te hniques or simple trial and error.

589
 If high-order ontrollers are designed for optimal performan e, ontroller
transfer fun tions with not h lter hara teristi s and/or high gains in small
frequen y bands are often the result. Su h a ontroller an deteriorate abruptly
in o-design onditions. Be ause of this reason, it is important that high-order
HI and MS ontrollers are not operated outside the range of their dened plant
un ertainties. If ontrollability of the air raft an be lost, the probability has
to be less than 10 9 o urren es per ight hour.
MF-25 splits requirements into a ategory that an be dealt with by feedfor-
ward ontrol and a ategory that is handled by feedba k ontrol. As robustness
is related to feedba k, robustness ae ts only the se ond ategory, whi h eases
the design task signi antly.
Robustness he ks by a worst- ase analysis or a -analysis test, whi h would
allow judgement of the a hieved stability performan e of all ontrollers, was not
available for the evaluators. The results of a -analysis test are dis ussed in
hapter 35.

34.2.5 Dis ussion and Lessons Learned

The Robust Flight Control design hallenge has been a unique han e to ben h-
mark new design te hniques. The high number of RCAM design entries (thir-
teen) ree ts the high interest of universities and resear h establishments in
ight ontrol problems. Mu h eort was invested into the denition of the
RCAM problem formulation, the RCAM model and the evaluation pro edure
 with a onsiderable result.

RCAM Problem Denition

The air raft model is a simplied model of a ivil transport air raft that has
well-behaved natural dynami s. In parti ular, the aerodynami model, engine
model, a tuator models and sensor models are simplied. The design mis-
sion, an automati ILS approa h without are, represents a small but relevant
sequen e of a typi al transport air raft mission. Air raft handling is simpli-
ed, i.e. no slat, ap, and landing gear extension. Operational onditions are
moderate: no rosswind on the nal approa h, moderate levels of turbulen e
and wind shear. No failures are onsidered, ex ept an engine failure during a
non- riti al ight phase. A lot of performan e, safety, passenger omfort and
ontrol surfa e requirements have been dened, representing a typi al extra t
of industrial requirements. More demanding are the robustness requirements
that ontrollers should be robust against: variations in mass, enter-of-gravity
lo ation, time delay, speed variations and engine failure. The design task is
mu h more realisti than typi al a ademi ight ontrol problems. A design,
from the study of the problem denition through to a presentable design report,
will require between three and nine months, depending on the experien e and
tools of the designer. The eort to make the ontroller ight-worthy is mu h
higher.

590
Evaluation Questionnaire  RCAM Experien e
The evaluation questionnaire proved to be an ee tive tool for evaluating the
dierent design methods and ontroller stru tures in a fair and omprehensive
manner. The missing results of the assessment software turned out to be an
advantage, as the evaluators fo used on the method and the ontroller stru -
ture. Some minor improvements should be made to the questionnaire in order
to avoid misunderstanding if it is to be used again.

RCAM Design and Evaluation

The teams have demonstrated modern design methods that emphasize robust
designs. The hoi e of synthesis methods was unrestri ted, resulting in a on-
entration around EA and HI/MS. This aused dupli ation of eort, repetition
of results and the impression of missing balan e. Other robust methods, for
example, quantitative feedba k theory (QFT), are not overed. In the future,
the appli ation of other methods should be en ouraged.
The evaluation results of ten design teams are dis ussed in detail above. All
designs have a hieved some level of su ess. But before any of the ontrollers
an be onsidered for implementation into a ight ontrol system, more detailed
evaluations are needed (investigation of disturban es, investigation of system
failure ases, system safety analysis, real-time ight simulation, et .).
The evaluation shows learly, that a lot of expert knowledge is needed for
ight ontrol design, not only in relation to the method, but also and foremost,
in relation to the pro ess (here: ight me hani s) and the requirements (here:
ying qualities and ight mission). Design methods that rely only on good on-
trol te hniques annot gain satisfa tory results. On the other hand, designers
that repeat traditional approa hes will be ome unable to introdu e real new
ideas. Modern methods have the potential to lead to new approa hes.
Finally, it should be said, that the RCAM design hallenge has ontributed
substantially to the progress of robust ight ontrol design. The GARTEUR
A tion Group has brought together the ompeten e from industry, resear h es-
tablishments and universities from all over Europe. The ollaboration proved
to be bene ial for all parti ipants. Industry was provided with a lot of useful
information on robust ontrol te hniques. Resear hers be ame aware of indus-
trial requirements on robust ight ontroller design and beyond. The dialogue
has started promising ollaboration whi h should be ontinued. The RCAM
ben hmark problem is a lasting out ome of the a tion group.

34.2.6 Re ommendations for Future Work

With thirteen designs performed, the RCAM model an be ome the ben hmark
problem for demonstrating new robust ight ontrol design methods to the ivil
air raft manufa turers. The RCAM model should be improved, taking into
a ount the lessons learned. Realisti but more demanding missions should
be dened in order to allow powerful methods a better demonstration of their

591
potential. Where possible, su h improvements should aim to keep ompatibility
with the existing work.
The RCAM model an be used to ben hmark other mu h more futuristi 
design te hniques, e.g. arti ial intelligen e methods su h as neural networks
and geneti algorithms, that have emerged quite re ently. These te hniques
have been applied with some su ess to omplex, poorly understood problems.
Although ight ontrollers are normally designed with high understanding of
ight dynami s, these novel methods might have some usable potential for the
treatment of failure onditions, where ight dynami s may hange dramati ally.
The presented ontinuous-time designs need an extra design step: the dis-
retization of the ontroller algorithm. Robust methods that dire tly design
dis rete ontroller algorithms would be very favourable, espe ially if high-order
ontrollers are the out ome (MS and HI). Resear h in this area is en ouraged.
It an be on luded that e onomi stresses are for ing the industry to redu e
ost and time to market. Engineers are asked to design more qui kly, ever more
omplex systems. This will stay as a running demand, as industry restru tures
and automates design pro esses to make them more e ient. Even design
methods for ight ontrollers are ae ted by su h pra ti al demands. Methods
and te hniques, su h as MO, that support an e ient automated ontrol design
pro ess, are needed and should attra t resear hers' attention.

A knowledgement
The author wants to thank the RCAM evaluation team:
W. Alles (DASA), R. Bro khaus (DLR), J. Breeman (NLR), J. Bos (NLR),
C. Fielding (BAe-MA), N. Foster (BAe-D), P. van der Geest (NLR), C. van
Gelder (NLR), G. Game (BAe-D), G. Grbel (DLR), M. Hut hinson (AVRO),
H.D. Joos (DLR), M. Horton (BAe-D), D. Laidlaw (AVRO), J.F. Magni (CERT),
A. Nieuwpoort (NLR), M. S hifaudo (ALN), D. Vorley (BAe-D), K. Weise
(DASA), J. Winter (DASA).
Spe ial thanks are due to Prof. Bro khaus who evaluated all design entries.

34.3 Military Air raft Manufa turer's View 12

34.3.1 Introdu tion

From an industrial viewpoint, in designing military air raft, the overall require-
ment is to be able to design a vehi le whi h meets the ustomer's operational
performan e requirements, is safe to operate and whi h an be designed and
built at an a eptable ost, within an agreed times ale. As an essential part
to a hieving the air raft's performan e and safety targets, the ight ontrol
system needs to be arefully designed, taking into a ount the requirements
and onstraints that will be imposed through airframe and system hardware
physi al limitations. The air raft's ight ontrol laws provide the basis for
a hieving the desired performan e hara teristi s and oer great potential for
12
Prepared by Chris Fielding on behalf of the HIRM Evaluation Team

592
operational exibility in terms of the possible pilot-sele table and automati
modes.
At an early stage of the ight ontrol laws design pro ess, primarily based on
experien e from other proje ts, it is possible to ease the later implementation,
testing and ight learan e aspe ts, by making appropriate assumptions. This
will involve hoosing the best ar hite ture for the ontrol laws, even before
any design te hnique or pro edure has been onsidered. It is essential to use
feedba k signals of su ient performan e and integrity, with multiplexing of
signals in order to a hieve overall system integrity targets. Additionally, ba k-
up modes may need to be designed to over for omplete loss of any sensor
information.
On e the ontrol laws ar hite ture has been established, the gains, lters
and nonlinear fun tions have to be designed. There are a wide range of (mainly
linear) te hniques for determining the appropriate parameter values. It is the
aim of the HIRM design hallenge and evaluation to identify the strengths and
weaknesses of a range of su h te hniques, in order to highlight the methods
whi h are best suited to support the ight ontrol laws design pro ess and
whi h are apable of produ ing a robust design.

34.3.2 Industrial Considerations for Appli ation of De-

sign Te hniques
In pra ti e, it is usual to design ight ontrol systems to be apable of provid-
ing good air raft handling qualities over a wide range of operating onditions
and to over the arriage of a range of internal and/or external stores. Addi-
tionally, new moding (e.g. autopilot modes) might be added to the baseline
system at a later date. At any stage of the design pro ess, a ompetent engi-
neer, other than the original designer, might inherit design responsibility and
must be able to take over the design without undue di ulty. From a wider
perspe tive, it must also be re ognised that there is a range of spe ialists and
managers who need to work with the ontrol laws at ea h stage of the total
design pro ess [123. These might be simulation engineers and pilots, ight
ontrol omputer implementers and testers, and ight learan e authorities.
The level of visibility of the fun tionality is important to these people, to help
them to gain su ient understanding to arry out their tasks.
As an essential part of the ontrol laws, the ontroller algorithms need to
be exe uted in real time in the air raft's ight ontrol omputer. Sin e this
apability is limited, it is important that ontrol algorithms are e ient and
do not lead to real-time pro essing problems. High order ontrollers, multi-
dimensional look-up tables and ompli ated nonlinear fun tions all add to om-
puter throughput requirements. There are also other aspe ts whi h ould ause
implementation di ulties. For example, the ontrol algorithm's numeri al a -
ura y requirements and its potential for gain s heduling, are all important
with respe t to implementation in the air raft's ight ontrol omputer.
The level of investment that industry will need to make in terms of devel-
oping design software to the appropriate quality standards and for training of

593
its designers to use a new method is also of interest; this needs to over the
ost of getting into a position of being able to use (and if ne essary, modify)
the method, su h that satisfa tory results an be a hieved, leading through a
straightforward ight learan e, to a su essful ight test programme.

34.3.3 HIRM Evaluation Team

An evaluation team was established, based on experien ed industrial aerospa e
ight ontrol systems designers, who themselves might be regarded as poten-
tial  ustomers for the range of design te hniques available. The industrial
ommunity have a genuine interest in the methods in terms of redu ed design
y le-time and asso iated osts, and improved produ ts. The use of industri-
alists on the evaluation team also supported a major obje tive of the urrent
exer ise, in transferring the knowledge and experien e from the GARTEUR
A tion Group into industry.
The evaluation team was led by British Aerospa e Military Air raft and
omprised experien ed ight ontrol law designers from the following organisa-
tions:

 British Aerospa e Military Air raft

 SAAB Military Air raft

 British Aerospa e Dynami s

Colle tively, this team has great experien e in the overall design of ight
ontrol systems and is highly apable of addressing all aspe ts of the design
in relation to the ight ontrol laws design pro ess. Additionally, the evalua-
tion team has olle tive experien e with appli ation of most of the proposed
methods, pla ing it in a strong position to perform the evaluation. Ea h or-
ganisation was provided with the HIRM design reports and its ight ontrol
spe ialists then evaluated the design entries and ompleted a set of question-
naires (as dened in the earlier se tion). The olle tive views and main points
whi h were made are dis ussed below, for ea h of the design entries.

34.3.4 Evaluation of the Design Entries

Linear Quadrati Approa h (LQ-26)
This design entry (Chapter 28) fo used on the Linear Quadrati approa h. This
well-established method is onsidered to be relatively easy to use, but requires
a good understanding of optimal ontrol theory. In terms of the learning urve
for ight ontrol appli ations, it was judged to be relatively easy to grasp.
Obtaining the desired results by sele tion of weighting fun tions will probably
provide the greatest di ulty for a new user. Some iteration may be ne essary
if the frequen y response hara teristi s are not satisfa tory.
The methodology results in a ontrol law ar hite ture similar to that for
a lassi al design, but whi h has a multivariable stru ture. Being essentially

594
a time-domain method, the transient response hara teristi s an be tuned to
some extent. The frequen y response hara teristi s have to be determined as
a se ond step and therefore there is likely to be some iteration in the design
(although this ould be automated).
The resulting ontroller is of low order, although for this parti ular design,
it was desirable to redu e the number of integrators. Implementation of the
ontroller should not be a problem, from the information provided and sys-
tem quali ation should be straightforward. With the proposed ar hite ture,
there is high potential for integrators winding up, espe ially sin e some of the
ontrolled variables are related through kinemati relationships. Although an
anti-wind-up te hnique is proposed for the parti ular ontrol s heme, the ve
integrators used in the design ould still give erti ation problems; this is a
feature of the ar hite ture hosen and not the design te hnique itself.
The ontrol law ar hite ture also ontains an undesirable swit hing logi
based on the pilot's roll ommand. This was an attempt to ompensate for
the ee ts of the gravity ve tor, during rolling manoeuvres. In pra ti e, the re-
quired ompensation an be expli itly in luded in the ontrol laws ar hite ture,
prior to designing any ontrollers.
The time histories presented were generally satisfa tory but there are low
stability levels indi ated for some ases.
Overall this was a very good eort, taking into a ount the limited ba k-
ground of the authors in ight ontrol problems.

Nonlinear Dynami Inversion and LQG (DI-30)

This design entry (Chapter 32) presents an interesting ombination of the Non-
linear Dynami Inversion and Linear Quadrati methods. Being su h a om-
bination, it obviously needs a greater level of understanding than ea h of the
methods individually. The most di ult part would appear to be asso iated
with the inversion of the air raft equations of motion, although on e estab-
lished, these equations would reside in the ontrol law framework. The method
does not appear to involve any di ult mathemati s.
The approa h onsidered handles the time response requirements well, by
shaping the feedforward ommands and then using feedba k to provide tight
tra king of these ommands (essentially a model following type of approa h).
Robustness requirements are not addressed dire tly but are assessed following
design, possibly leading to design iterations. The robustness aspe ts ould be
addressed by introdu ing an automated pro edure.
The results presented indi ate good robustness and performan e, with tight
tra king of the feedforward terms. Control a tivity is as to be expe ted, but
some spikey responses and rate limiting were noted. The high frequen y gains
in the system (from inertial sensors through to a tuation ommands were not
visible in the design as presented  noting that the HIRM design hallenge on-
strains these gains to ensure a pra ti al design and avoid undue airframe/FCS
aeroservoelasti ity problems).
In general, this was another good design entry whi h in the NDI part of the

595
design made good use of ight me hani s knowledge.

Robust Inverse Dynami s Estimation Approa h (DI-28)

This design entry (Chapter 33) addresses the Robust Inverse Dynami s Esti-
mation approa h to ontrol laws design and ontains relatively straightforward
mathemati s. The method indire tly attempts to provide the appropriate air-
raft for es and moments to trim the air raft and de ouple the outputs, while
proportional plus integral loops provide the desired set of air raft response
hara teristi s. The method is related to Nonlinear Dynami Inversion.
The method gives perfe t results for the nominal model in the absen e of
hardware dynami s; for this relatively trivial ase, the nominal design is very
easy to a hieve. As hardware models and data un ertainty are introdu ed, more
eort will be needed and iterations required for tuning the ontroller to meet
robustness riteria. The eort needed will depend on the degree to whi h the
nominal design is representative of the full high-order modelling assumptions,
in luding the un ertainties.
The method is good for shaping the time histories but does not handle
robustness requirements dire tly (it is similar to Eigenstru ture Assignment in
this respe t). Although robustness is measured after the initial ontroller has
been designed and this may result in manual tuning, it should be possible to
automate the iteration pro ess, to in lude robustness riteria.
In terms of the stru ture of the design a hieved, a short oming is asso i-
ated with the ontroller's high frequen y end-to-end gains whi h should have
been onstrained to meet the HIRM's design spe i ation, in order to pro-
vide a realisti level of attenuation at airframe stru tural mode frequen ies.
For implementation purposes the omplexity of the design is onsidered to be
a eptable, with a ontroller of low order.
The Ni hols plots presented indi ate the design to be robust (it is essen-
tially three de oupled single loop designs). The ontroller's performan e also
appears to be satisfa tory, with the time histories meeting their spe ied values
for bandwidth and handling qualities riteria. Overall ontrol surfa e a tivity
seems to be a eptable.
This design entry provided an interesting and intuitive approa h whi h was
easy to follow due to the links to ight me hani s whi h were taken into a ount
in the ontrol law ar hite ture. The method is relatively new and oers poten-
tial for rapid prototyping of initial designs. Further work needs to be arried
out to develop the method for pra ti al appli ations, in luding implementation.

H1 , -Synthesis Approa h (MS-29)

This design entry (Chapter 31) presented a very detailed appli ation of the
-synthesis approa h. In order to use this approa h for ight ontrol laws
design, a thorough understanding of the theory is needed to ee tively design
and analyse ontrollers. The method is onsidered to be relatively di ult to
grasp, due to the mathemati s involved and any ontrol engineer without a
strong mathemati al ba kground is likely to struggle to learn this method.

596
 The design formulation presented allowed almost all of the design require-
ments to be in luded. Good use was made of the relationships between air raft
physi al parameters in relation to the time domain handling qualities metri s.
In ommon with some of the other design entries, the gravity ve tor terms
seemed to have aused undue problems in the design. These should be in luded
before designing the ontrollers , in order to provide some feedforward to
help to minimise sideslip during rolling manoeuvres. Additionally the absen e
of velo ity ve tor rolling meant that the the - ontroller inherently in luded
an equivalent term to suppress sideslip, and may have resulted in higher gains
than ne essary. The lateral design was somewhat unusual in that it did not
in lude a yaw rate feedba k term.
The resulting ontroller's total order, with 35 states (longitudinal and lateral-
/dire tional), is far higher than might be expe ted for this type of ontrol prob-
lem (e.g. a lassi al design would be about 10th order). It is not lear what
benets this large number of additional states oer, in terms of improved ro-
bustness and/or performan e. Real time exe ution is likely to be a problem
and therefore further order redu tion is onsidered to be essential.
In terms of erti ation, the method results in bla k box ontrollers whi h
la k the visibility whi h is essential for ight erti ation purposes. For this
parti ular design entry, the ontroller had some visibility, in that tables of
ontroller eigenvalues and an array of ontroller frequen y response plots were
in luded (the original design had some eigenvalues whi h ould not be realised
within the HIRM's assumed 80 Hertz digital omputing design onstraint; these
were later designed out). A further impli ation of the high order ontrollers is
that gain s heduling to over a wider envelope may also lead to timing di ul-
ties for real-time implementation, although it is a knowledged that pro essing
power is always improving and therefore, although this is a valid on ern it
may not be justied.
The robustness of the design appears to be good, whi h is to be expe ted
for this method. It is not lear how mu h of the a hieved robustness is due
to the high order of the ontroller and what would be lost if the order were
redu ed. The performan e of the design also looks to be good, with indi ations
of good handling qualities.
The ontrol surfa e dee tions used for manoeuvring are as might be ex-
pe ted but the symmetri al anard, rudder and throttle RMS values due to
moderate turbulen e seem to be high (note that a ben hmark value is not
available for omparison purposes).
Overall this was a very good eort, with a high level of omplian e with all
aspe ts of the HIRM design hallenge do umentation.

A -Synthesis Approa h (MS-31)

This was a further design entry (Chapter 30) on the -synthesis loop shaping
approa h and not surprisingly, some of the ndings from the evaluations were
therefore similar to those dis ussed above.
As noted from the previous entry, the designer would need to have an up-to-

597
date familiarity with the mathemati s involved, in order to be able to gain an
understanding of the te hnique. For the design there is a step-by-step pro ess,
although a detailed understanding appears to be ne essary for su essful sele -
tion of weighting fun tions. It should however, be possible for engineers who
an work in terms of magnitudes and frequen ies to arry out a -synthesis
design, without knowing the internal details of the method.
For this method to gain a eptan e from pra tising ight ontrol engineers,
a visible and intuitive explanation of the te hnique is essential. The method
was judged to be a bla k box approa h, whi h would undoubtedly ause
di ulties from a erti ation point of view due to its la k of visibility. Some
visibility would be gained if the state-spa e ontrollers were split into lters,
giving the transfer fun tions from the dierent feedba k signals to the ontroller
output signals.
Also in ommon with the previous entry, the design, for what is a fairly
straightforward problem, has resulted in a signi antly higher order ontroller
than would be expe ted using a more  onventional approa h  although it
does redu e the requirement for air data s heduling (at least for the ight
envelope onsidered). The issue of the di ulties in gain s heduling of the high
order ontrollers was also raised during evaluation. As the results were for an
an initial design, it is expe ted that the order of the ontroller ould be redu ed
by using model redu tion te hniques.
The results a hieved demonstrate satisfa tory hara teristi s for most ases
but learly showed that the xed ontrollers would have benetted from gain
s heduling, to normalise the ee ts of dynami pressure.
Overall, it was onsidered that this was another good design entry with a
high emphasis on the ight ontrol aspe ts, with a well balan ed set of on-
lusions. The strong point of this entry was the manner in whi h the HIRM
design problem was interfa ed with the -synthesis pro edure and in parti -
ular, how the design requirements for robustness and handling qualities were
a ommodated.

H1 Loop Shaping Approa h (HI-32)

This was a further entry (Chapter 29) in the  H1 / ategory, whi h serves
to emphasise the urrent level of interest in this methodology. As with the two
previous entries, some ommon ndings were made during the evaluations.
It was again found that some of the mathemati s was too detailed for in-
dustrial ight ontrol engineers and ould not be easily understood by them,
unless they had undertaken signi ant resear h in this area (or perhaps, had
re ently graduated with a Masters degree). However, the authors laim that a
detailed understanding of the mathemati s is not a pre-requisite for applying
the method.
It was evident that the method involved a good read-a ross from the single-
input single-output prin iples asso iated with lassi al ontrol. The user does
need to understand singular values quite thoroughly though, in luding their
physi al interpretation. It was judged that this method is more di ult to

598
grasp than some of the other te hniques available. Some interesting alternative
metri s for measuring robustness were introdu ed.
The method overs all frequen y domain requirements but annot handle
time domain requirements dire tly, although this was not seen to be a problem,
sin e these an usually be a hieved by lassi al ommand path lter shaping.
An alternative approa h is to onvert the time domain handling requirements
into frequen y domain spe i ations.
The ontroller has a favourable level of visibility in terms of its overall ar-
hite ture but this be omes lost within the state-spa e blo ks. As a result,
it is not possible to he k the ontrollers hara teristi s without resorting to
omputer-based te hniques. If the resulting high order ontrollers have to be
s heduled with several parameters, there will be a signi ant omputing re-
quirement. The la k of the state-spa e blo k's visibility is likely to lead to
additional erti ation issues.
This method should a hieve the most robust ontroller possible (although
other methods might possibly have a hieved a similar design). From the few
time histories presented, the performan e looks good, but the bandwidths have
been maximised and there is no visibility that the design onstraint on the
allowable feedba k gains had been satised. Indeed, some of the time histories
show signi ant rate-limiting and spikes in the ontrol surfa e responses, whi h
is indi ative of ex essive high frequen y gains.
This was a good eort overall, with some good rationale and assumptions
made for handling the ight ontrol aspe ts. This helps the reader understand
the ar hite ture hosen.

34.3.5 Dis ussion and lessons learned

HIRM Design Challenge: Problem Denition
It is rstly important to note that the HIRM design hallenge poses a rela-
tively straightforward task, ompared with urrent industrial design problems,
parti ularly for air raft with high levels of inherent instability (the HIRM is ap-
proximately neutrally stable). Additionally, the HIRM being ee tively a low
Ma h number subsoni air raft, did not require onsideration of air ompress-
ibility ee ts. However, it is onsidered that given the limited time available
to the design teams, the problem as dened was suitable for its purpose. In-
deed, the HIRM design hallenge has posed a more realisti problem than that
whi h might usually be onsidered for ight ontrol resear h and has provided
a good introdu tion to the ight ontrol design problem for the organisations
whi h had no previous involvement in this area; it has given them a basis, and
hopefully the motivation, for arrying out further work on ight ontrol.
The HIRM problem denition was probably too ompli ated in some areas
and la king enough detail in others. In parti ular, a wider ight envelope would
have enfor ed gain s heduling to show the true omplexity of some of the on-
trollers proposed. The air raft model should have in luded the ee ts of moving
the air raft's entre of gravity fore and aft, sin e this would have provided a
good robustness test with an extremely simple physi al interpretation.

599
 The HIRM design problem was, to some extent, an unknown element for
the designers and the evaluators. It was based on a DRA drop model high
in iden e aerodynami dataset, augmented with representative FCS hardware
assumptions and design riteria provided by BAe-MA. Sin e the vehi le was
not a known air raft, no ben hmark design existed, against whi h relative
merits of alternative designs ould be assessed. With hindsight, it would have
been preferable to have based the exer ise on a known air raft.

Evaluation Questionnaire  HIRM Experien e

The use of the questionnaire for extra ting relevant information was su essful,
sin e mu h information was gathered in a onsistent format. This made eval-
uation relatively easy and provided a good basis for the information presented
in this hapter. From the responses to some of the questions, the questionnaire
ould have been improved slightly; it would have been bene ial to arry out
more detailed advan e testing of the questionnaire, had proje t times ales been
less demanding.

HIRM Design Challenge: Evaluation of Results

Overall, the results and omments produ ed by independent evaluations showed
good agreement; without wishing to repeat the details from the previous se -
tions, some more general points regarding the results were made by the evalua-
tors. At the time when the HIRM Evaluation Team was a tive, the quantitative
Automati Evaluation results were not available and any omparisons made
are therefore, mainly of a qualitative nature.
Very good agreement was a hieved for the level of understanding and learn-
ing urve for su essful appli ation of the methods, the handling of design
requirements by the methods and the resulting design omplexity. All evalua-
tors, in luding the designers (as self-evaluators), had good experien e in these
areas.
Greater variations in evaluation results were seen for estimated re-design
eort, ontrol law stru tures whi h ould be a ommodated by the method,
ontroller visibility and ontroller implementation. The variations in ratings
were mainly asso iated with variations in experien e of the evaluators and
designers. Greatest variation in evaluation results was seen for ight learan e
and quali ation aspe ts, where the evaluators' experien e varied onsiderably
and the designers' experien e was minimal.
Overall, the balan e of the design entries was heavily on entrated around
H1 / and this led to some dupli ation of eort and repetition of results. With
hindsight, it would have been preferable to limit the number of design entries
in ea h ategory, to en ourage a wider assessment of the te hniques available.
It was learly established, that for ee tive design of ight ontrol laws,
the designer needs to have, as an absolute minimum, a good understanding
of ight me hani s and ight ontrol, in order to establish a suitable ontrol
laws ar hite ture. An inappropriate layout will inevitably lead to a poor design
and the benets due to using any design te hnique are likely to be devalued

600
or even lost in the pro ess. The ontroller design te hnique itself was seen as
a small, but signi ant part of the total design pro ess, when we take into
a ount all the nonlinear aspe ts of ight ontrol whi h have to be addressed.
Some of the designs suered from a la k of knowledge of ight me hani s and
in parti ular, gravity ve tor ompensation, velo ity ve tor rolling and dynami
pressure ee ts were not always in luded. Additional signi ant fa tors whi h
required due onsideration are: inertial oupling, integrator onditioning, rate
limiting and angle of atta k and airframe loading limiting fun tions.
For all the methods it was judged that any minor re-design should be
straightforward, provided that established tools, weighting fun tions and/or
inter onne tion stru ture ould be used. However, for any major re-design,
perhaps due to a signi ant hange in air raft hara teristi s, re-design would
be more di ult; for example, if new weighting fun tions have to be deter-
mined. The level of eort required will, in general, be parti ular to the design
problem.
In pra ti e, robustness has to be demonstrated for all possibilities, in order
to a hieve ight erti ation. For the so- alled bla k box ontrollers this was
seen to be a potential problem. Additionally, it was noted that the urrent
pro edures are based on past and urrent methods and would probably need
to be developed in order to be ompatible with some of the proposed methods.
All the designs whi h have been produ ed for the HIRM have a hieved some
level of su ess, but from the results available, more work would be needed to
develop these to a ight-worthy standard; this to be expe ted from an initial
design. Further work would also be required to enable relative levels of a hieve-
ment to be reliably judged  ea h design would need to be fully optimised
with respe t to the design riteria. The evaluation exer ise would have benet-
ted from a -analysis for omparing the robustness of the ontrollers produ ed,
parti ularly for the designs with multiple feedba k loops.
Although the robustness tests that were arried out were onsidered to be
adequate, a -analysis test would have been simple and onsistent; the other
tests would have been used as supporting information, to aid interpretation.
To a hieve the obje tives of the robust ight ontrol design hallenge, it
was of vital importan e that results showing good performan e were presented.
Despite a good design eort by very apable design teams, some of the results
were onsidered to be of an a ademi  nature and were not onvin ing enough
for the aeronauti al industry, and tended to raise many unanswered questions.
In many ases, the visibility of what had really been a hieved did not be ome
apparent until the nonlinear time histories were presented and in some ases,
these did not always support the linear robustness results (possibly due to rate
limiting ee ts). Finally, in fairness to the design teams, they did not really
have the time to arry out a omplete design of ontrol laws and some were
starting with a very limited knowledge of ight ontrol.
As noted earlier in des ribing the evaluation pro ess, the su essful design
of ontrol laws is dependent on many fa tors and owes mu h to the designers,
their working environment and the pro edures used to provide a framework for
the design method used. It is onsidered that, given the right onditions, most

601
(or even all) the methods an be used su essfully. The results from the HIRM
design hallenge have onrmed that the methods an all be made to work,
with ea h method having its own strengths and weaknesses.
The la k of a ben hmark design for the HIRM did not allow any benets
from the various design methodologies to be quantied. However, from the
method des riptions and the results presented, the benets (over a lassi al
design approa h) are asso iated with establishing a logi al framework for man-
aging the linear design, whi h expli itly in ludes both design spe i ations and
modelling un ertainties. A further advantage is oered by using optimisation
algorithms to establish an e ient route to a hieving a design. Alternative mea-
sures of robustness were proposed and these should be bene ial over lassi al
measures, provided they an be alibrated in terms of physi al hara teristi s.
Disadvantages observed were asso iated with design visibility and omplexity,
parti ularly if the ontrollers were of ex essively high order and presented in
state-spa e format. It is onsidered that these aspe ts ould be addressed by
improved presentation and do umentation of the ontrollers to provide total
visibility and by further developments in model redu tion te hniques to redu e
omplexity.

General Dis ussion

The ight ontrol resear h ommunity that is developing and applying new
methods to air raft models is relatively large, ompared to the numbers of in-
dustrial pra titioners or resear h s ientists a tively designing ontrol laws for
implementation and ight testing. Arguably there are many design te hniques
hasing relatively few real-life ight ontrol appli ations. It is re ognised how-
ever, that the te hniques have potential value for an extremely wide range of
ontrol engineering appli ations and development should therefore be en our-
aged; ight ontrol provides a relevant, interesting and hallenging platform for
the development of su h te hniques.
If a new design te hnique is to be a epted by any industrial ight ontrol
systems design ommunity, it is essential that it an be made to be ompatible
with their existing design pro ess (whi h itself is ontinually being reviewed and
developed) and that pra ti al aspe ts, su h as implementation in the air raft's
ight ontrol omputer, have not been ignored. If the new method annot be
a ommodated easily, then there will inevitably be low interest and a natural
relu tan e to adopt the method.
From the individual designer's viewpoint, he will tend to use the method
with whi h he feels omfortable and an a hieve his obje tives, within the
available times ale. Usually, this results in the designer using the method
with whi h he has most familiarity and any hange in approa h is naturally
di ult (perhaps seen as a risk), parti ularly if a high level of su ess has been
previously a hieved. For the individual, a hange an only really be justied if
a satisfa tory design is not possible with his existing approa h (although this
ould, of ourse, be due to design problem physi al limitations) or if a qui ker
and/or more ost-ee tive means of a hieving an a eptably similar design an

602
be demonstrated.
For an organisation, there may be dierent reasons to hange to a dierent
method su h as, the lower level of skill needed to use some of the newer methods
due to their inherent automation (for example, brought about by optimisation).
For su h methods, the designers need to be aware that any automation, whi h
has been introdu ed into the design pro ess, potentially removes them from the
design loop, possibly at the expense of losing some design insight (e.g. they
might lose the feel for any design trade-os, if the omputer has performed this
a tivity for them). Irrespe tive of the method, su h automation needs to be
introdu ed arefully, in order to provide the designer with a summary of what
route has been taken to a hieve the design, and the results at ea h stage (i.e.
an audit trail). Engineering skills must not be sa ri ed in the pro ess; it is
essential that the ontrol law designer retains authority and that this is not
given to the omputer and design tools used for the design.
An aim of the GARTEUR A tion Group was to provide greater awareness
in industry of the use of the new methods. Although this has been a hieved,
it has been a two-way ex hange, with mu h useful information owing from
industry ba k into the external resear h ommunity. This has highlighted some
of the pra ti al aspe ts that industry has to onsider, beyond robustness and
performan e. This should enable the resear hers to expand their urrent design
drivers, to onsider the pra ti alities of implementation and ight erti ation.
Quite learly, ollaboration and ex hange of information is the best way to
fo us method developments, if industrial appli ation is the target.
Ultimately, the only way for an industrial organisation to truly determine
the benets of any method in relation to their design problems, is to arry out
a detailed assessment using suitably experien ed personnel. Su h hands-on
experien e will allow the benets to be evaluated in a ontrolled environment.
The HIRM design hallenge has indi ated what might be expe ted from the
te hniques and should en ourage greater interest in the range of methods avail-
able.

34.3.6 Re ommendations for future work

For ontrol methods development, it is important that the benets, whi h an
be attributed to the use of a parti ular design te hnique, are demonstrated
and quantied in a fair and meaningful manner. Therefore resear hers need
to make every eort to demonstrate the benets of the new te hniques, in
order to promote their methodology within industry. It is essential that good
visibility and physi al links are established to allow theoreti al measures to be
applied and related to ight ontrol problems, in order to provide engineering
insight and results that are relevant and easily interpreted by the aeronauti al
engineer. In general, the status of many of the available tools needs to be
elevated from resear h tools to engineering tools whi h interfa e with the
industrial design pro ess and whi h meet the appropriate software standards.
In terms of spe i developments of the methods, it is onsidered that fur-
ther developments in the following areas would be of benet:

603
  Controller visualisation te hniques.

 Controller order redu tion te hniques.

 Means of measuring design method benets.

 Development and alibration of robustness measures.

 Nonlinear analysis development (e.g. to over rate limiting ee ts).

To omplement the methods development work, industry needs to prepare

further hallenging and realisti ben hmark problems, preferably based on ex-
isting air raft, to enable the ight ontrol resear hers to address genuine indus-
trial appli ations. There needs to be more dis ussion between resear h design
teams and industry to ensure that design requirements and onstraints are fully
understood and in luded in design studies.
A follow-on GARTEUR a tivity on the ight ontrol subje t, possibly using
the HIRM, should be onsidered. The team whi h has been established on the
urrent exer ise, has shown that it an work well together with a high level
of ommitment. Industry has proposed some interesting and relevant topi s
for possible further studies. A GARTEUR exploratory group now needs to be
established, in order to determine the level of interest in greater detail.

A knowledgement
This hapter would not have been possible without the support of the HIRM
Evaluation Team. Design evaluations were arried out by the following per-
sonnel: Mike Walker (BAe-MA), Jonathan Irving (BAe-MA), Steve Barratt
(BAe-MA), Karin Sthl-Gunnarsson (SAAB-MA), Robert Hillgren (SAAB-
MA), Per-Olov Elg rona (SAAB-MA), Kenneth Eriksson (SAAB-MA), Lars
Rundquist (SAAB-MA) and George Game (BAe-D).
Finally, the work of three other HIRM design teams is a knowledged  led
by Alex Smerlas (ULES) for HI-33, Aldo Tonon (Alenia) for CC-24 and Jrgen
A kermann (DLR) for MO-27.

604
35. Another View on the Design
Challenge A hievements

Georg Grbel 1

Abstra t: The value of the Design Challenge as a atalyst for

further advan ing robust ontrol theory, and for stratifying the on-
trol design pro ess, is reviewed. Based on the RCAM robust ontrol
design-entries data base, a need is re ognized for improved robust
synthesis tuning, and for a post-design stability-robustness assess-
ment based on linear (LFT) system models, whi h are automati ally
generated from parameterized nonlinear air raft dynami s models.

35.1 Introdu tion

A valuable a hievement of the GARTEUR Design Challenge on Robust Flight
Control is a oherent data base of air raft ight ontrol designs, gained by
dierent design teams using dierent design methods. Although this represents
no omplete overage of state-of-the-art ontrol design methods, - notably the
QFT method [116 is missing -, the ompiled results and the experien e gained
in the design pro ess, allow some general ndings by asking:

 Is ight ontrol relevant to robust ontrol theory? Meaning, does ight

ontrol impose a hallenge in further developments of robust ontrol the-
ory, algorithms, and software tools?

 Is robust ontrol theory relevant to ight ontrol design methodology?

Meaning, does robust ontrol theory stratify a 'best-pra ti e' design pro-
ess?

A view on these two aspe ts may be seen as a methods-oriented evaluation

of the Design Challenge a hievements, whi h is omplementary to the industrial
view of the previous hapter. The ndings are partly based on an independent
post-design stability robustness assessment of the nal RCAM design entries.
However, it is believed that the same arguments hold for the HIRM ben hmark
as well.

1
DLR German Aerospa e Resear h Establishment, Institute for Roboti s and System
Dynami s, Control Design Engineering, D-82234 Wessling, E-mail: georg.gruebeldlr.de

605
35.2 Relevan e to Robust Control Theory
A ording to industrial evaluation, all nal design entries ould handle, - better
or worse -, the posed ight- ontrol ben hmark requirements.
The design entries dier in their ontrol stru tures. They use either a
'ight-me hani s dened' ontrol stru ture, or an 'analyti ally dened' on-
trol stru ture (e.g. state feedba k), or a 'pilot-dened' nonlinear-gain feed-
ba k stru ture whi h results from a rule base of desired ight behaviour within
a fuzzy-logi ontrol approa h. The design methods dier in the way, how
parameter un ertainties are modelled in the design synthesis pro ess: either
parameter un ertainty is dire tly modelled as a parameter interval (MS-19,
2
LY-14) , or parameter un ertainty is dealt with in a multi-model approa h
(MO-16, EA-12), or the ee ts of parameter un ertainties are redu ed to plant
gain- and phase intervals or to high-frequen y onstraints on sensitivity fun -
tions (MS-11, EA-22, HI-21), or there is no un ertainty modelling at all (FL-15,
CC-13, MF-25).
If a ontrol design method is hara terized to be a robust ontrol design
method by the fa t that synthesis model un ertainties are expli itly dealt with
in the ontrol-synthesis formalism, some methods (FL-15, CC-13, and MF-25)
do not qualify as 'robust' methods, although they have been su essfully applied
in the Design Challenge. On the other hand, expli it un ertainty modelling in
a suited synthesis formalism does not lead to a robust design per se, without
a skilful synthesis-parameter tuning. This was experien ed by some design
teams.
Flight ontrol laws are quite omplex. For reasons of ' ontroller visibility',
they are usually designed sequentially for longitudinal/lateral and inner/outer
loop ontrol, possibly applying dierent ontrol stru tures and dierent synthe-
sis methods for the dierent ontroller loops. For example, a lassi al inner-loop
feedba k has been ombined with a fuzzy-logi outer-loop design. Or, a state-
feedba k longitudinal ontroller has been ombined with a 'ight-me hani s de-
ned' lateral ontroller with wash-out lters et . Hen e, un ertainty modelling
whi h is used in a parti ular synthesis step, may not be adequate and general
enough for an overall robustness assessment. Rather, a synthesis-independent
'post-design' stability-robustness assessment is required. Flight ontrol, in this
respe t, poses a hallenge for robust ontrol theory be ause of the rather om-
plex operational deviations and dynami s un ertainties to be dealt with.
'Robust ontrol tools' [32 are needed to readily apply numeri ally e-
ient robustness analysis te hniques. Parameterized LFT (Linear Fra tional
Transformation) models are a general system-des ription form for des ribing
parameter-un ertain linear system models, and thereby an be used as a stan-
dard form for whi h su h tools should be appli able. Be ause air raft dynami s
models are highly parameterized, there is a need for an automated generation of
LFT models from the given physi al un ertainty des ription. Proper CACSD
tools (see next se tion) an redu e this tedious task from days to minutes; but
there remain theoreti al and omputational issues to be solved, for example

2
abbreviations see Table 35.1

606
the automated derivation of minimum blo k-size LFT des riptions.

An independent post-design robustness assessment has been performed [158

for the RCAM ben hmark, having just the ontroller Matlab S-fun tions of ten
design entries available no matter how they were synthesised. For all (lineariz-
able) ontrollers and over all un ertainty parameters, two approa hes of robust-
ness analysis have been arried out. A  stru tured singular value robustness
analysis was performed for the given nominal airspeed. This approa h with
symboli linearization of the air raft equations did not allow to over airspeed
deviations properly. Therefore, in addition, a worst- ase un ertainty-parameter
sear h for eigenvalue relative stability (minimum damping of all eigenvalues)
has been performed within the whole range of the ight envelope. The worst-
ase eigenvalue analysis was omputed by a numeri al optimization loop over
the un ertain parameters, involving numeri al trimming of the air raft and
system linearization. The results are ompiled in Table 35.1.

Table 35.1 shows, that not all ontrollers are stability-robust for all possible
un ertainty-parameter ombinations within the ight envelope. In part, this is
due to the fa t that this post-design robustness analysis has been performed for
a larger deviation in verti al enter-of-gravity position
3 and also for airspeed
4
deviations , whi h both had not been spe ied as a design requirement. Hen e
the designers must not be blamed for any orresponding ontroller de ien ies.
But, most results in Table 35.1 demonstrate, that good overall robustness is
indeed possible.

Table 35.1 shows that low-order ontrollers may exhibit better stability-
robustness properties than ontrollers with higher-order dynami s. This is
mainly a question of tuning the synthesis parameters (e.g. weighting oe-
ients and weighting lters, or ontroller parameters) in order to make best use
of the intrinsi apabilities of the orresponding synthesis method, or the ho-
sen ontroller stru ture. Hen e, systemati tuning methods for robust ontrol
synthesis should be developed. Su h synthesis-tuning methods are likely to be
based on intera tive numeri al optimization su h as the Methodof Inequalities
(MOI) [255, or the method of Multi-Obje tive Parameter Synthesis (MOPS) 2
used in 16. The general need for optimization-based synthesis tuning in air raft
ight ontrol design is also re ognized in [253.

Sin e ontroller-order redu tion is part of the synthesis pro edure, 'opti-
mal' tuning is also required for this task. Synthesizing the lowest-order ro-
bust ontroller for a given appli ation is a question of multi-obje tive perfor-
man e/robustness ompromising. It is still an issue in air raft ight ontrol,
due to the limited throughput of the available Ele troni Flight Control System
omputers. This hallenge was not su iently emphasized by most of the de-
sign teams. The large range of a hieved ontroller-dynami s order is interesting
to note in Table 35.1.

3
A modelling error in dealing with the verti al enter-of-gravity position was dis overed
too late to be taken are of properly by the designers.
4
The requirement to he k robustness w.r.t. deviations from nominal airspeed has been
added late in the proje t.

607
35.3 Relevan e to the Control Design Pro ess
Control design is an (iterative) a tivity within the triangle of the plant physi al
system, design goals, and the algorithmi ontroller apabilities. This is visu-
alized by Figure 35.1. The required ontrol theory for the inner design-triangle
is supported by CACSD (Computer Aided Control System Design) algorithms,
tools, and frameworks, e.g. [97. Commonly available toolboxes support (lin-
ear) plant analysis and ontrol synthesis. Online data-base support is required
for a transparent and intera tive logging of a multi- riteria set up, whi h is
the formal link between design goals and the analysis and synthesis a tivities.
The CACSD design triangle is substantiated by ight- ontrol system evalua-
tion and validation. This is supported by nonlinear simulation experimenting,
in luding rapid prototyping with iron-bird and in-ight simulation, as well as
hardware-in-the-loop man/ma hine interfa e validation by using o kpit ight-
simulators.

DESIGN GOALS

mathematical modelling of
performance & robustness
eva

multi
ers

lua

criteria
tun

ting

set-up
sis

dyn
the

am
syn

ics

CACSD
ting

ind
ap

ica
ad

tor
s

control plant
synthesis analysis
control code (linear) synthesis model physical system- &
generation uncertainty modelling
simulation experimenting
nonlinear closed-loop system models
rapid prototyping H/W-in-the loop

ALGOR. CONTROLLER PHYSICAL PLANT

Figure 35.1: The ontrol-design inner and outer 'a tivity triangle'.

Commonly, ontrol-dynami s design is seen as an a tivity on two strata,

namely (linear) plant analysis/ ontrol synthesis, and design validation by (non-
linear) simulation. Robust ontrol theory makes a third design stratum evident.
Mathemati al modelling of robustness riteria in addition to mathemati ally de-
ned performan e riteria, leads to a multi- riteria problem, formalized by a
multi- riteria set up, see Figure 35.1. Using the stratum of a multi- riteria

608
set up allows omputer-aided synthesis tuning on the basis of an expli it om-
parison of a hieved results against a set of design requirements. This is more
systemati than manual tuning, having multi- riteria requirements in one's
mind only. In 16 it is shown how a multi- riteria set up an be dire tly used
for adapting synthesis tuners in an automated way. It is also shown there, how
most of the design goals an be mapped onto 're-usable' standard mathemati al
riteria des riptions to be evaluated via standard dynami s indi ators, su h as
eigenvalues, time responses, and frequen y responses.
The basi requirement of robust ontrol, namely, that real-world design
un ertainties are to be expli itly taken into a ount, yields the requirement that
the inner design-triangle of Figure 35.1 is to be expli itly linked to the outer
'real-world' triangle. This impa ts the ight ontrol design pro ess by the need
for modelling both the physi s and the un ertainty behaviour. For both kinds of
modelling, the use of standard forms is advisable as a 'best-pra ti e approa h'
be ause the 're-usability' of standard forms makes standardized simulation-,
analysis- and synthesis tools readily appli able.
Two kinds of standard model forms for des ribing a parameter-un ertain
physi al system are demonstrated in the design hallenge. These are, (i), re-
usable omponent lass libraries for nonlinear, physi ally parameterized obje t-
oriented air raft-system dynami s modelling [145, 177, and, (ii), linear, phys-
i ally parameterized LFT models whi h are (semi-)automati ally derived from
(i). The basis for that is symboli /numeri pro essing of omponent-equation
type system models.

bodyfixed experimental airspeed u atmos

body6DOF bodyLong RCAM
kinetic
Trafo Trafo
air
COG COG
veh.carried bodyfixed wind aeroRCAM engine

HIRM const. gravity 1dim gravity const. 1D

Earth Earth
Engine wind gust atmosphere atmos1D

Figure 35.2: The RCAM/HIRM air raft dynami s omponent library [145, 177
for graphi al 'pi k & drag' system-dynami s aggregation.

The usefulness of formalized symboli system modelling, Figure 35.3, is

demonstrated by the fa t, that any nonlinear ight dynami s simulation ode
used by the RCAM and HIRM Design Challenge teams (i.e. Matlab mex ode
for Simulink and neutral DSblo k Fortran- and C- ode) was derived by auto-
mati ode generation from a 'single-sour e' symboli model, whi h itself is soft
programmed in an intera tive 'pi k and drag' system modelling environment
using respe tive omponent- lass libraries, Figure 35.2.
From the symboli nonlinear air raft dynami s model, a standard LFT
model an be automati ally derived [245 via symboli linearization and us-

609
 graphical pick & drag system aggregation

physical system model component class libraries

mathematical
Matlab / Simulink
(symbolic) model
automatically generated S-function (cmex)
from physical system
objects aggregation &
class libraries via neutral DSblock
modelling environment Simulation - model
Dymola Fortran-, C-Code

symbolic linearisation LFT - model

.
via Maple
[ x = A (p) x + B (p) u ]
automatically generated
by PUM

Figure 35.3: Formalized symboli /numeri system modelling for parameterized

system analysis and simulation.

ing the Matlab toolbox PUM (Parametri Un ertainty Modelling) [239. This
was demonstrated for the RCAM ben hmark and served for the independent
post-design  analysis as referred to in the previous se tion. Sin e an LFT-
model is a general type of model des ription, a system-theoreti onne tion of
two LFT system models is again an LFT system model. This allows LFT mod-
els derived from signal-ow diagrams [261 to be ombined with LFT models
derived from parameterized physi al models as des ribed above.
In on lusion, robust ontrol theory per se, does not yield basi ally new air-
raft ight ontrol system ar hite tures. In fa t, as demonstrated in the Design
Challenge, su ient experien e in ight me hani s and ight ontrol allows you
to design quite robust ontrol laws on the basis of any 'method you know best',
- provided that enough engineering time is spent for proper robustness tuning
and re-iterative post-design robustness analysis. The required engineering time
for robustness tuning and post-design robustness analysis, however, an be sig-
ni antly redu ed by a omputer-automated design pro ess whi h is stratied
taking into a ount the basi issues of robust ontrol. Essentially, these are
the multi- riteria nature of synthesis-parameter tuning, and the availability of
numeri ally reliable algorithms for stability robustness analysis.

610
 type- dynami  V = Vnom min = V 
entry no. order up bnd unstable min nom unstable
due to: due to:
FL-15 5 (nl) 0.44
EA-12 9 0.36 0.27 0.47
EA-22 9 0.39 0.15 0.23
CC-13 11 0.51 0.04 0.17
MO-16 12 0.35 0.21 0.52
MS-19 35 1.36 Tdel
1) -0.13 -0.26
1)
MF-25 36 0.65 0.05 0.17
HI-21 36 1.53
2) -0.18 -0.98
2)
LY-14 39 0.57 0.11 0.19
MS-11 62 0.49 -0.05 -0.27 v
3)

Type: FL = Fuzzy Logi , MM = Modal Multi-Model Synthesis,

EA = Eigenstru ture Assignment,
CC = Classi al Control, MO = MOPS, MS =  Synthesis,
MF = Model Following, HI = H 1 , LY = Lyapunov

Changing parameters: [m; delx; delz; Tdel; v

mass: m 100 000 kg  m  150 000 kg
.o.g.-x: delx 0.15  delx  0.31
.o.g.-z: delz 0.0  delz  0.21
time delay: Tdel 0.05s  Tdel  0.1s (pad approx.)
air speed: v 1.23 V (m)Stall  V  90 m/s
Vnom = 80 m/s
gain/phase 0  k
k1  0.01
added in 
( analysis)
instability parameters:
1) [100; 0.15; 0.21; 0.1; 90.0
2) [150; 0.31; 0.21; 0.1; 71.8
3) [100; 0.15; 0.21; 0.1; 58.2

Table 35.1: Results of RCAM independent post-design robustness assessment.

611
36. Con luding Remarks

Samir Bennani , Jean-Franois Magni and Jan Terlouw

1 2 3

The main purpose of the Design Challenge was to demonstrate, by means of

design and omputer simulations, how twelve advan ed te hniques an be ap-
plied to design robust ontrollers for two fairly realisti ight ontrol ben hmark
problems. The hosen level of omplexity of the RCAM and HIRM problem
denitions was a trade-o between industrial relevan y and feasibility for the
design teams to nd solutions in a limited period of time. In order not to
hamper the teams in showing the potentialities of the applied methods, the
idea was to leave them as mu h freedom as possible: freedom in the trade-o
between a hieving good robustness and performan e and freedom in the hoi e
of the ar hite ture. This has led to designs whi h speak for themselves and we
have not ompared the results here in any detail. We believe that this book
has signi ant tutorial value, while the two evaluation hapters give a broad
and dierentiated view on the potential of the methods.

To some extent, the Design Challenge has proven that modern te hniques
an be used to design ontrollers for realisti problems. Additionally, it has on-
rmed that requirements for industrial appli ation of new te hniques are quite
severe. From an industrial point of view, desirable features of any te hnique an
be assumed to be: transparen y, simpli ity, quality, a ura y, delity, reliabil-
ity, implementability, predi tability and generality. Even though the presented
methods have mu h potential in the eld of improved robustness, better per-
forman e, de- oupled ontrol and simpli ation of the design pro ess, some of
them do not yet have the maturity required for industrialisation. Even ma-
ture methods need to be arefully integrated into the industrial design pro ess
to fully address the omplexities asso iated with modern air raft. One of the
main problems en ountered remains the omplexity of the proposed ontrol
solutions, whi h is partly driven by the hoi e of the ontrol ar hite ture. This
is a ru ial a tivity in the design pro ess, whi h is not yet taken into a ount
su iently by the theoreti al ommunity.

1
Delft University of Te hnology, Fa ulty of Aerospa e Engineering, Stability and Control
Group, Kluyverweg 1, 2629 HS Delft, The Netherlands, E-mail: s.bennanilr.tudelft.nl
2
CERT ONERA, Dpartement d'tudes et Re her hes en Automatique, BP 4025, F31055
Toulouse Cedex, Fran e
3
National Aerospa e Laboratory NLR, Flight Me hani s Department, Anthony Fokker-
weg 2, 1059 CM Amsterdam, The Netherlands.

612
Future resear h
Several software tools and do uments have been developed by GARTEUR A -
tion Group FM(AG08) and used su essfully throughout the proje t:

 Manuals des ribing the RCAM and HIRM ben hmark denitions, in lud-
ing the air raft dynami s models, the environmental models (atmospheri
turbulen e, windshear, et .), a traje tory generator for RCAM, the design
requirements, an evaluation questionnaire and guidelines for the prepa-
ration of design reports.

 Matlab/Simulink RCAM and HIRM design environments and evaluation

tools, with whi h designs an be ompared with the requirements.

 A ommon nomen lature.

 Twenty-one RCAM and HIRM design reports, whi h ontain more details
than the design hapters of Parts II and III.

To en ourage broader parti ipation in the Design Challenge problems, the

RCAM manual [145, the a ompanying software and most of the design re-
ports are made available and an be obtained from the NLR homepage
(http:\\www.nlr.nl).
Information about HIRM an be found at this omputer lo ation as well.
The information listed above an be used for edu ational, resear h and
appli ation purposes at universities, resear h establishments and in industry.
It may serve as an introdu tion for people who are interested in how parti ular
advan ed ontrol te hniques work. Furthermore, in general, the ben hmark
problems an be used for a riti al review of design methods.
Improvements and extensions of the ben hmarks are within rea h. For
HIRM, for example, it would be interesting to allow a variation in the lo ation
of the entre of gravity to introdu e relaxed air raft stability. Design spe i-
 ations for both problems an be further improved and it may be advisable
to propose ontroller ar hite tures for a next Design Challenge. This would
ease the omparison of design methods. Extensions of the ight envelopes for
RCAM and HIRM would make gain s heduling ne essary in most ases. It
would be interesting to see how advan ed te hniques ope with this additional
omplexity.
It is likely that in Europe aeronauti al resear h and te hnology development
proje ts will be arried out more and more on a multi-national s ale. The
work arried out by GARTEUR A tion Group FM(AG08) has benetted from
two positive aspe ts. Firstly, results were obtained in only two years be ause
twenty-three organisations worked very well together, in parallel but as a team.
Se ondly, the lose ollaboration between many people from industry and the
s ienti ommunity seems to have a elerated the pro ess of bringing ( ontrol)
theory to pra ti e. It is re ommended to ombine these two aspe ts again in
future European resear h proje ts.

613
Appendix A. Used Nomen lature

Anders Helmersson and Karin Sthl Gunnarsson

1 2

A.1 Introdu tion

This hapter denes the re ommended nomen lature to be used within FM(AG
08).
In the hapter, nomen lature regarding oordinate frames, transformations
between oordinate frames, variables des ribing the air raft as well as some
mathemati al quantities is proposed.
Generally, ve tor notation has not been used. The user of the hapter is
free to dene ve tor notation if that is preferred.
The alphanumeri des riptors are aimed to apply in software. Throughout
this hapter apital letters are given, but lower ase letters an used if preferred.
If the software is restri ted to only have short variable names, for instan e
with only six letters, as in F77 standard Fortran, the proposed alphanumeri
des riptors an be shortened in a suitable way.
ISO units shall be used in al ulations, while industrial units, su h as de-
grees, Hz and dB, an given in presentations.

A.2 Denition of Referen e Frames

This se tion deals with the denition of referen e frames. The referen e frames
that are dened are the earth-xed referen e frame, the vehi le- arried verti-
al frame, the body-xed referen e frame, the wind axis referen e frame, the
stability axes referen e frame and the measurement referen e frame.
All referen e frames are right-handed orthogonal.

A.2.1 Earth-Fixed Referen e Frame FE , (OE XE YE ZE )

The Earth-xed referen e frame is dened a ording to Figure A.1.
FE is an earth-surfa e frame, with origin OE near the vehi le. ZE is positive
downward, XE is positive north and YE is positive east.

A.2.2 Vehi le-Carried Verti al Frame FV , (OV XV YV ZV )

The vehi le- arried verti al frame is dened a ording to Figure A.2.

1
Linkping University, S-581 83 Linkping, Sweden
2
Saab Military Air raft, S-581 88 Linkping, Sweden

614
 YE

XE

OE

ZE

Figure A.1: The earth-xed referen e frame

YV (East)

OR OV
(North)

XV C.G.
g

Ref. point
aerodata
ZV

Figure A.2: The vehi le- arried verti al frame

This frame is moving with the vehi le and is parallel to the earth-xed
frame. The origin OV is atta hed to the vehi le at the enter of gravity. ZV is
positive verti ally downward along the g ve tor. XV is positive north and YV
is positive east.

A.2.3 Body Axis Referen e Frame FB (OB XB YB ZB )

The body axis referen e frame is dened a ording to Figure A.3.

The origin OB is at the vehi le's enter of gravity. XB is positive forward,

ZB is positive downward and YB is positive to the right.

For the transformation from vehi le arried verti al frame to body axes, the
vehi le- arried verti al frame is rst rotated by the heading angle . The next
rotation is by the pit h angle  and the last rotation is by the roll angle .

615
 YB

OR OB

C.G.

Ref. point
aerodata
XB ZB

Figure A.3: The body axis referen e frame

A.2.4 Wind Axis Referen e Frame FW ( OW XW YW ZW )

The wind axis referen e frame is dened a ording to Figure A.4. This frame
has the origin OW atta hed to the vehi le at the enter of gravity. XW is
positive pointing forward along the velo ity ve tor V. ZW is positive downward,
and YW is positive to the right.
For the transformation from body axis referen e frame to wind axis referen e
frame, the body axis system is rst rotated by the angle  and then by the
angle .
The denitions of angle of atta k, , and sideslip,  , are:
w
 = ar tan a
ua
v
 = ar sin a
VA
where ua , va and wa are the x-, y- and z- omponents of the ve tor airspeed Va
expressed in body axes, and VA is total velo ity
p
VA = (u2a + va2 + wa2 ):

A.2.5 Stability Axis Referen e Frame FS (OS XS YS ZS )

The stability axis referen e frame is dened a ording to Figure A.5. The
originOS of this frame is at the enter of gravity of the vehi le. XS is positive
forward, ZS is positive downward and YS is positive to the right.
For the transformation from body axes to stability axes, the body axis
system is rotated by the angle .

616
 YW YB

XB

Ref. point
OR aerodata

XW OB O
V W

C.G.

ZB

ZW

Figure A.4: The wind axis referen e frame

A.2.6 Measurement Referen e Frame FM (OM XM YM ZM )

The measurement referen e frame is dened a ording to Figure A.6. The
origin OM of this frame is at the leading edge of the mean aerodynami hord.
XM is positive pointing ba kwards, YM is positive pointing to the right and
ZM is positive pointing up.

A.3 Coordinate Transformations

The transformations asso iated with single rotations about the three oordinate
axes as shown in Figure A.7 are now dened.

2 3
1 0 0
R1 (X1 ) = 4 0 os X1 sin X1 5
0 sin X1 os X1

2 3
os X2 0 sin X2
R2 (X2 ) = 4 0 1 0 5
sin X2 0 os X2

2 3
os X3 sin X3 0
R3 (X3 ) = 4 sin X3 os X3 0 5
0 0 1

617
 YB Y S

XB

Ref. point
OR aerodata

XS OB O
S

V
C.G.

ZB

ZS

Figure A.5: The stability axis referen e frame

ZM

XM
YM

OM c OR

Ref. point
aerodata

Figure A.6: The measurement referen e frame

A.3.1 Vehi le- arried verti al frame to body axes

The transformation matrix RBV onsists of the sequen e of rotations
RBV = R1 ()R2 ()R3 ( )
The result of multiplying the three matri es is
2 3
os  os os  sin sin 
RBV = 4sin  sin  os os  sin sin  sin  sin + os  os sin  os 5
os  sin  os +sin  sin os  sin  sin sin  os os  os 

A.3.2 Body Axis to Vehi le-Carried Verti al Frame

The transformation matrix RVB onsists of the sequen e of rotations
RVB = R3 ( )R2 ( )R1 ( )

618
The result of multiplying the three matri es is

2 3
os  os os sin  sin  sin os sin sin  + os sin  os 
RVB = 4sin os  os os  +sin sin  sin  sin sin  os  os sin 5
sin  os  sin  os  os
Note that
T .
RV B = RBV

A.3.3 Body Axes to Wind Axes

The transformation matrix RW B onsists of the sequen e of rotations

RW B = R3 ( )R2 ( )
The resulting matrix is

2 3
os  os  sin  sin  os 
RW B = 4 os  sin  os  sin  sin  5
sin  0 os 

A.3.4 Body Axes to Stability Axes

The transformation matrix RSB onsists of the sequen e of rotations

RSB = R2 ( )
The resulting matrix is

2 3
os  0 sin 
RSB = 4 0 1 0 5
sin  0 os 

A.3.5 Wind Axes to Body Axes

The transformation matrix RBW onsists of the sequen e of rotations

RBW = R2 ()R3 (  )
The resulting matrix is

2 3
os  os  os  sin  sin 
RBW = 4 sin  os  0 5
sin  os  sin  sin  os 
T .
RBW = RW
Note that B

619
A.3.6 Stability Axes to Body Axes
The transformation matrix RBS onsists of the sequen e of rotations

RBS = R2 ()
The resulting matrix is
2 3
os  0 sin 
RBS = 4 0 1 0 5
sin  0 os 
Note that
T .
RBS = RSB
xa3 xb1 xa1
xb3

xb2 xb3
X1 X2
O xa2 O xa3

a) b)

xb2 xa2

xb1
X3
O xa1
c)

Rotation about a) xa1 b) xa2 c) xa3

Figure A.7: The basi rotations

A.3.7 Quaternion Attitude Representation

An alternative to the use of Euler angles for dening attitudes, is to use the
quaternion representation. This is a four- omponent entity whi h enjoys some
advantages ompared to the Euler angles: it is numeri ally e ient and it has
no singularities.

620
 The quaternion parameters

q = (q0 ; q1 ; q2 ; q3 )T
full the following ondition

q02 + q12 + q22 + q32 = 1

and are related to the Euler angles by

q0 = ( os(=2) os(=2) os( =2) + sin(=2) sin(=2) sin( =2))

q1 = (sin(=2) os(=2) os( =2) os(=2) sin(=2) sin( =2))
q2 = ( os(=2) sin(=2) os( =2) + sin(=2) os(=2) sin( =2))
q3 = ( os(=2) os(=2) sin( =2) sin(=2) sin(=2) os( =2))
Whi hever sign is hosen in the above equations must be applied onsistently
a ross all equations.
The equations above an be used to al ulate the initial values of the quater-
nions parameters given the initial values of the Euler angles.
The derivatives of the quaternions with respe t to time an be expressed as

2 3 2 32 3
q_0 0 p q r q0
6 q_1
6
7 1
7= 6
6p 0 r q 7 6 q1
76
7
7
4 q_2 5 2 4 q r 0 p 5 4 q2 5
q_3 r q p 0 q3
where p, q and r are the body-axes angular rates.

621
A.4 Aeronauti al Variable Symbols and
Alphanumeri Des riptors
A.4.1 Air raft Related Quantities
Name Symbol Alphanumeri Unit
Wingspan b B m
Mean aerodynami hord  CBAR m
Air raft inertia tensor I IAC kgm2
x body axis moment of inertia Ix IX kgm2
x-y body axis produ t of inertia Ixy IXY kgm2
x-z body axis produ t of inertia Ixz IXZ kgm2
y body axis moment of inertia Iy IY kgm2
y-z body axis produ t of inertia Iyz IYZ kgm2
z body axis moment of inertia Iz IZ kgm2
Radius of gyration, x dire tion in FB rx RX m
Radius of gyration, y dire tion in FB ry RY m
Radius of gyration, z dire tion in FB rz RZ m
Centre of gravity x position in FB X g XCG m
Centre of gravity y position in FB Y g YCG m
Centre of gravity z position in FB Z g ZCG m
Lo ation of aerodynami entre x position Xa XAC m
in FB
Lo ation of aerodynami entre y position Ya YAC m
in FB
Lo ation of aerodynami entre z position Za ZAC m
in FB
Generalized length l LEN m
Air raft total mass m MASS kg
Wing planform area S S m2
Vehi le weight W W N
Tail unit planform area St STAIL m2
Longitudinal distan e between the lt LTAIL m
aerodynami entre of the wing and body
and the aerodynami entre of the tail unit
Displa ement of aerodynami entre
r DELR m
from entre of gravity
Displa ement of aerodynami entre
x DELX m
from entre of gravity along x body axis
Displa ement of aerodynami entre
y DELY m
from entre of gravity along y body axis
Displa ement of aerodynami entre
z DELZ m
from entre of gravity along z body axis

622
Name Symbol Alphanumeri Unit
Downwash angle  EPSILON rad
Angle from the thrust axis of engine to T EPSILONT rad
the x-y body axis plane
Angle between x body-axis and prin ipal p EPSILONP rad
axis of inertia
Angle from the proje tion of engine thrust  XSI rad
ve tor (Fp ) onto the engine x-y plane to
the lo al x-axis
Angle from engine thrust ve tor ( Fp ) to  ETA rad
the engine x-y plane

A.4.2 Engine Quantities

Name Symbol Alphanumeri Unit
X position of entre of gravity of XCGE1 XCGE1 m
engine 1 in FB
Y position of entre of gravity of YCGE1 YCGE1 m
engine 1 in FB
Z position of entre of gravity of ZCGE1 ZCGE1 m
engine 1 in FB
Engine thrust ve tor at entre of gravity Fp FP N
X position of point of appli ation of thrust XAT P 1 XATP1 m
of engine 1 w.r.t. entre of gravity in FB
Y position of point of appli ation of thrust YAT P 1 YATP1 m
of engine 1 w.r.t. entre of gravity in FB
Z position of point of appli ation of thrust ZAT P 1 ZATP1 m
of engine 1 w.r.t. entre of gravity in FB
Rotational inertia of the engine Ie IE kgm2
Angular momentum of engine rotor he HE kgm2 =s
Mass of engine mE MASSE kg
Moment from engine about the TE TE Nm
entre of gravity
Gyros opi torque from engine TEg TEG Nm
Engine angular velo ity !E OMEGAE rad/s

623
A.4.3 Air Data Quantities
Name Symbol Alphanumeri Unit
A eleration due to gravity g G m=s2
g = 9:80665 m2 =s
Speed of sound in air a A m/s
Pressure p PRESSURE N=m2
Ambient pressure pa PA N=m2
Total pressure pt PT N=m2
Dynami pressure q QBAR N=m2
Impa t pressure q QC N=m2
Reynolds number Re RE -
Reynolds number per unit length Re0 REPRIME 1/m
Ambient temperature T TAMB K
Total temperature Tt TTOT K
Coe ient of dynami vis osity  CVISC kg=ms
Density of air  RHO kg=m3
Airspeed ve tor in FB Va VAVEC m/s
x- oordinate of Va in FB ua UA m/s
y- oordinate of Va in FB va VA m/s
z- oordinate of Va in FB wa WA m/s
p
Total airspeed VA = (u2a + va2 + wa2 ) VA VAIR m/s
Calibrated airspeed V VC knots
True airspeed VT AS VTAS knots
Equivalent airspeed Ve VEAS knots
Ma h number M MACH -
S ale of turbulen e Lu ; Lv ; Lw LU, LV, LW m
Mean square value of gust velo ity u SIGMAU m
v SIGMAV m
w SIGMAW m

A.4.4 Energy Related Quantities

Name Symbol Alphanumeri Unit
Spe i energy Es ES m
Spe i power Ps PS m/s

624
A.4.5 Variables
Name Symbol Alphanumeri Unit
Altitude (relative to mean sea level, h = z) h H m
Roll rate p P rad/s
Pit h rate q Q rad/s
Yaw rate r R rad/s
Total inertial velo ity V V m/s
Inertial velo ity in FB x dire tion u UB m/s
Inertial velo ity in FB y dire tion v VB m/s
Inertial velo ity in FB z dire tion w WB m/s
Angle of atta k  ALPHA rad
Angle of sideslip  BETA rad
Pit h angle  THETA rad
Roll angle  PHI rad
Heading angle PSI rad
Bank angle (=aerodynami roll angle)  BANK rad
x position of entre of gravity in FE x X m
y position of entre of gravity in FE y Y m
z position of entre of gravity in FE z Z m
Quaternion parameter q0 Q0 -
Quaternion parameter q1 Q1 -
Quaternion parameter q2 Q2 -
Quaternion parameter q3 Q3 -
Velo ity of wind in FE x dire tion WXE WXE m/s
Velo ity of wind in FE y dire tion WY E WYE m/s
Velo ity of wind in FE z dire tion WZE WZE m/s
Velo ity of wind in FB x dire tion WXB WXE m/s
Velo ity of wind in FB y dire tion WY B WYE m/s
Velo ity of wind in FB z dire tion WZB WZE m/s

625
A.4.6 Derivatives of Variables
Name Symbol Alphanumeri Unit

Time rate of hange of altitude h_ HDOT m/s

Time rate of hange of roll rate p_ PDOT rad=s2
Time rate of hange of pit h rate q_ QDOT rad=s2
Time rate of hange of yaw rate r_ RDOT rad=s2
Time rate of hange of total inertial V_ VDOT m=s2
velo ity
Time rate of hange of FB x dire tion u_ UBDOT m=s2
velo ity
Time rate of hange of FB y dire tion v_ VBDOT m=s2
velo ity
Time rate of hange of FB z dire tion w_ WBDOT m=s2
velo ity
Time rate of hange of angle of atta k _ ALPHADOT rad/s
Time rate of hange of angle of sideslip _ BETADOT rad/s
Time rate of hange of pit h angle _ THETADOT rad/s
Time rate of hange of roll angle _ PHIDOT rad/s
Time rate of hange of heading angle _ PSIDOT rad/s
Time rate of hange of bank angle _ BANKDOT rad/s
Time rate of hange of x position of x_ XDOT m/s
entre of gravity in FE
Time rate of hange of y position of y_ YDOT m/s
entre of gravity in FE
Time rate of hange of z position of z_ ZDOT m/s
entre of gravity in FE
Time rate of hange of quaternion q_0 Q0DOT -
Time rate of hange of quaternion q_1 Q1DOT -
Time rate of hange of quaternion q_2 Q2DOT -
Time rate of hange of quaternion q_3 Q3DOT -

A.4.7 Flight Path Related Quantities

Name Symbol Alphanumeri Unit
Flight path angle GAMMA rad
Time rate of hange of ight path angle _ GAMMADOT rad/s
Flight path a eleration  GAMMADDOT rad=s2
Tra k angle  CHI rad
Time rate of hange of tra k angle _ CHIDOT rad/s
Verti al a eleration h HDDOT m=s2

626
A.4.8 Control Surfa e Dee tions
Name Symbol Alphanumeri Unit
Aileron dee tion A DA rad
Dierential anard dee tion CD DCD rad
Symmetri anard dee tion CS DCS rad
Elevator dee tion E DE rad
Rudder dee tion R DR rad
Speed brake dee tion SB DSB rad
Tailplane dee tion T DT rad
Dierential taileron dee tion T D DTD rad
Symmetri taileron dee tion T S DTS rad
Throttle position of engine 1 T H 1 THROTTLE1 -
Throttle position of engine 2 T H 1 THROTTLE2 -
Su tion su tion SUCTION -

A.4.9 A elerations and Load Fa tors

Name Symbol Alphanumeri Unit
Normal a eleration at entre of gravity an AN m=s2
in FB
Normal a eleration, not at entre of ANI m=s2
gravity, in FB an;i
x body axis a elerometer output at anx ANX m=s2
entre of gravity, in FB
y body axis a elerometer output at any ANY m=s2
entre of gravity, in FB
z body axis a elerometer output at anz ANZ m=s2
entre of gravity, in FB
x body axis a elerometer output, not anx;i ANXI m=s2
at entre of gravity, in FB
y body axis a elerometer output, not any;i ANYI m=s2
at entre of gravity, in FB
z body axis a elerometer output, not anz;i ANZI m=s2
at entre of gravity, in FB
Load fa tor in FB , n = mgZ n N g
Fx n
FB , nx = mg
Longitudinal load fa tor in x NX g

Lateral load fa tor in FB , ny =

Fy n
mg y NY g

Normal load fa tor in FB , nz =

Fz n
mg z NZ g

627
A.4.10 For e Quantities
Name Symbol Alphanumeri Unit
Drag for e, positive ba kward, in FW D DRAG N
Side for e, positive right, inFW YW YWFORCE N
Lift for e, positive upward, in fW L LIFT N
Longitudinal for e, positive ba kward, T TFORCE N
in FB
Lateral for e, positive right, in FB Y YFORCE N
Normal for e, positive upward, in FB N NFORCE N

A.4.11 Body Axis Quantities

Name Symbol Alphanumeri Unit
A eleration along the x-axis ax AX m=s2
A eleration along the y-axis ay AY m=s2
A eleration along the z-axis az AZ m=s2
Total for e along the x-axis Fx FX N
Total for e along the y-axis Fy FY N
Total for e along the z-axis Fz FZ N
Aerodynami for e along the x-axis FxA FXA N
Aerodynami for e along the y-axis FyA FYA N
Aerodynami for e along the z-axis FzA FZA N
Thrust along the x-axis FxT FXT N
Thrust along the y-axis FyT FYT N
Thrust along the z-axis FzT FZT N
In remental x-axis for e X DFX N
In remental y-axis for e Y DFY N
In remental z-axis for e Z DFZ N
Total rolling moment L LTOT Nm
Total pit hing moment M MTOT Nm
Total yawing moment N NTOT Nm
Aerodynami rolling moment LA LA Nm
Aerodynami pit hing moment MA MA Nm
Aerodynami yawing moment NA NA Nm
Thrust ontribution to rolling moment LT LT Nm
Thrust ontribution to pit hing moment MT MT Nm
Thrust ontribution to yawing moment NT NT Nm
In remental rolling moment L DL Nm
In remental pit hing moment M DM Nm
In remental yawing moment N DN Nm

628
A.4.12 Coe ients of For es and Moments
For the denition of oe ients like Cm_ , Cmq , Cl_ , et ., the following onven-
tions are used:

q  _ 
q = _ =
2V 2V
pb _
b
p = _ =
2V 2V
rb
r =
2V

Name Symbol Alphanumeri Unit

Coe ient of drag CD CD -
Coe ient of drag at zero angle of atta k CD 0 CD0 -
Coe ient of drag due to angle of atta k CD CDA -
Coe ient of drag due to angle of atta k CD_ CDAD -
rate
Coe ient of drag due to elevator CDE CDDE -
dee tion
Coe ient of drag due to symmetri CDCS CDDCS -
anard dee tion
Coe ient of drag due to symmetri CDT S CDDTS -
taileron dee tion
Coe ient of drag due to speed brake CDSB CDSB -
dee tion
Coe ient of axial for e CX CX -
Coe ient of sidefor e CY CY -
Coe ient of normal for e CZ CZ -
Coe ient of sidefor e due to angle of CY  CYB -
sideslip
Coe ient of sidefor e due to angle of CY _ CYBD -
sideslip rate
Coe ient of sidefor e due to roll rate CY p CYP -
Coe ient of sidefor e due to yaw rate CY r CYR -
Coe ient of sidefor e due to aileron C Y A CYDA -
dee tion
Coe ient of sidefor e due to rudder C Y R CYDR -
dee tion
Coe ient of sidefor e due to dierential CY CD CYDCD -
anard dee tion
Coe ient of sidefor e due to dierential C Y T D CYDTD -
taileron dee tion

629
Name Symbol Alphanumeri Unit
Coe ient of lift CL CLFT -
Coe ient of lift at zero angle of atta k CL0 CLFT0 -
Coe ient of lift due to angle of atta k CL CLFTA -
Coe ient of lift due to angle of atta k rate CL_ CLFTAD -
Coe ient of lift due to pit h rate CLq CLFTQ -
Coe ient of lift due to elevator dee tion CLE CLFTDE -
Coe ient of lift due to symmetri anard CLCS CLFTDCS -
dee tion
Coe ient of lift due to symmetri taileron CLT S CLFTDTS -
dee tion
Coe ient of lift due to speed brake CLSB CLFTDSB -
dee tion
Coe ient of longitudinal for e CT CT -
Coe ient of longitudinal for e at zero CT 0 CT0 -
angle of atta k
Coe ient of longitudinal for e due to CT  CTA -
angle of atta k
Coe ient of longitudinal for e due to CT _ CTAD -
angle of atta k rate
Coe ient of longitudinal for e due to CT q CTQ -
pit h rate
Coe ient of longitudinal for e due to CT E CTDE -
elevator dee tion
Coe ient of longitudinal for e due to CT CS CTDCS -
symmetri anard dee tion
Coe ient of longitudinal for e due to CT T S CTDTS -
symmetri taileron dee tion
Coe ient of longitudinal for e due to CT SB CTDSB -
speed brake dee tion
Coe ient of normal for e CN CN -
Coe ient of normal for e at zero angle CN 0 CN0 -
of atta k
Coe ient of normal for e due to angle CN CNA -
of atta k
Coe ient of normal for e due to angle CN _ CNAD -
of atta k rate
Coe ient of normal for e due to pit h rate CNq CNQ -
Coe ient of normal for e due to elevator CNE CNDE -
dee tion
Coe ient of normal for e due to CNCS CNDCS -
symmetri anard dee tion
Coe ient of normal for e due to CNT S CNDTS -
symmetri taileron dee tion

630
Name Symbol Alphanumeri Unit
Coe ient of normal for e due to speed CNSB CNDSB -
brake dee tion
Coe ient of rolling moment Cl CL -
Coe ient of rolling moment due to angle Cl CLB -
of sideslip
Coe ient of rolling moment due to anglee Cl_ CLBD -
of sideslip rat
Coe ient of rolling moment due to Clp CLP -
roll rate
Coe ient of rolling moment due to Clr CLR -
yaw rate
Coe ient of rolling moment due to ClA CLDA -
aileron dee tion
Coe ient of rolling moment due to ClCD CLDCD -
dierential anard dee tion
Coe ient of rolling moment due to ClT D CLDTD -
dierential taileron dee tion
Coe ient of rolling moment due to ClR CLDR -
rudder dee tion
Coe ient of pit hing moment Cm CM -
Coe ient of pit hing moment at zero Cm0 CM0 -
angle of atta k
Coe ient of pit hing moment due to Cm CMA -
angle of atta k
Coe ient of pit hing moment due to Cm_ CMAD -
angle of atta k rate
Coe ient of pit hing moment due to Cmq CMQ -
pit h rate
Coe ient of pit hing moment due to CmE CMDE -
elevator dee tion
Coe ient of pit hing moment due to CmCS CMDCS -
symmetri anard dee tions
Coe ient of pit hing moment due to CmT S CMDTS -
symmetri taileron dee tion
Coe ient of pit hing moment due to CmSB CMDSB -
speed brake dee tion

631
Name Symbol Alphanumeri Unit
Coe ient of yawing moment Cn CN -
Coe ient of yawing moment due to Cn CNB -
angle of sideslip
Coe ient of yawing moment due to angle Cn_ CNBD -
of sideslip rate
Coe ient of yawing moment due to Cnp CNP -
roll rate
Coe ient of yawing moment due to Cnr CNR -
yaw rate
Coe ient of yawing moment due to CnA CNDA -
aileron dee tion
Coe ient of yawing moment due to CnR CNDR -
rudder dee tion
Coe ient of yawing moment due to CnCD CNDCD -
dierential anard dee tion
Coe ient of yawing moment due to CnT D CNDTD -
dierential taileron dee tion

632
A.5 Mathemati al Quantities, Symbols and
Alphanumeri Des riptors

A.5.1 Matri es and Norms

Name Symbol Alphanumeri Unit

Unit matrix I EYE -
Unit matrix of dimension n In EYEN -
Square root of -1 j JAY -
Complex onjugate x XBAR -
Tranpose of omplex onjugate xH XH -
Transpose xT XTP -
Pseudo-inverse xy XPSINV -
The i; j element of matrix [X ij or xij XIJ -
Determinant det(X ) DETX -
Tra e tr(X ) TRX -
The ith eigenvalue i (X ) EIGXI -
Largest and max (X ) EIGXMIN -
smallest eigenvalue min (X ) EIGXMAX -
Perron-Frobenius eigenvalue p (X ) EIGPFX -
Stru tured singular value (G) SSVG -
Spe tral radius (X ) SRX -
The ith singular value i (X ) SVXI -
Largest and  SVXMAX -
smallest singular value  SVXMIN -

Name Symbol Alphanumeri Unit

1-norm, maxj i jxij j jjxjj1 NORM1X -
Frobenius norm jjxjjF NORMFX -
Spe tral norm jjxjjS NORMSX -
H2 -norm jjGjj2 NORMH2X -
Hankel norm jjGjjH NORMHANKX -
Innity norm jjGjj1 NORMINFX -
-norm jjGjj NORMMUX -
A sto hasti pro ess x(t) X -
Auto ovarian e fun tion xx ( ) COVX -
Power spe tral density xx(!) SPECX -
Krone ker delta jk KRONDELTAJK -

633
A.6 System Des riptive Symbols and
Alphanumeri Des riptors
A.6.1 System Des ription
Name Symbol Alphanumeri Unit
State matrix of the state equation A A -
Control matrix of the state equation Bu BU -
Disturban es matrix of the state equation Bw BW -
State matrix of the output equation C C -
Control matrix of the output equation Du DU -
Disturban es matrix of the output equa- Dw DW -
tion
State ve tor x X -
Derivative of state ve tor x_ XDOT -
Referen e signal r R -
Error signal e E -
Control signal u U -
Measurement error m M -
Measured variables y Y -
Exogenous input signal w W -
Regulated output signal z Z -
Transfer fun tion of plant G(s) G -
Transfer fun tion of ontroller K (s) K -
Transfer fun tion of pre- ompensator P (s) P -
Transfer fun tion of feedforward ompen- F (s) F -
sator of disturban es
Open-loop transfer fun tion from w to z Pzw PZW -
Open-loop transfer fun tion from u to z Pzu PZU -

634
Name Symbol Alphanumeri Unit
Open-loop transfer fun tion from w to y Pyw PYW -
Open-loop transfer fun tion from u to y Pyu PYU -
Closed-loop transfer fun tion from w to z Tzw TZW -
Return dieren e Fi (s), FI, FO -
Fo (s)
Sensitivity fun tion S (s) S -
Complementary sensitivity T (s) T -
Additive perturbation
a (s) DELTAA -
Input multipli ative perturbation
i (s) DELTAI -
Output multipli ative perturbation
o (s) DELTAO -
1
Normalized un ertainty, Wi
Wj
1 ~ s)

( DELTANORM -
Weighting fun tions W (s) W -
Crossover frequen y ! WC rad/s
Bandwidth ( 3 dB) !b WB rad/s
Zeros zi ZEROI -
Poles pi POLEI -
1 Am
Gain margin (
jGK180o j ) GM -

Phase margin m PM rad

Lapla e operator s S rad/s
Undamped natural frequen y !n WN rad/s
Damping ratio  ZETA -
Spatial frequen y
OMEGA rad/m
Cir ular frequen y ! WFREQ rad/s

A.6.2 Subs ripts

Name Symbol Alphanumeri Unit
Body axis B B -
Centre of gravity CG CG -
Earth-xed referen e frame E E -
Stability axis S S -
Vehi le- arried verti al frame V V -
Wind axis W W -
Measurement referen e frame M M -
Wing/body onguration wb WB -
Tail t T -
Command C -

635
Bibliography
[1 Spe ial issue on the LQG problem, February 1971. IEEE Transa tion on Au-
tomati Control, Vol AC-16.

[2 Military standard, ying qualities of piloted vehi les, MIL-STD-1797A (USAF),

31 Mar h 1987.

[3 Handling qualities of unstable highly augmented air raft, AGARD Advisory

Report No.279, AGARD AR-279, May 1991.

[4 GARTEUR guide, published by the GARTEUR Coun il. GARTEUR X/D-16,

Se retariat: NLR, PO Box 90502, NL-1006 BM Amsterdam, August 1993.

[5 S. Abu el Ata-Doss, P. Fiani, and J. Ri halet. Handling input and state on-
straints in predi tive fun tional ontrol. In Pro eedings, Control and De ision
Conferen e, Brighton UK. IEEE, 1991.

[6 R.C. Adams, J.M. Bungton, Sparks A.G., and Banda S.S. Robust multi-
variable ight ontrol, Advan es in Industrial Control. Springer-Verlag, Berlin,
1994.

[7 R.J. Adams, J.M. Bungton, and S.S. Banda. Design of nonlinear ontrol
laws for high-angle-of-atta k ight. AIAA Journal of Guidan e, Control, and
Dynami s, 17(4), 1994.

[8 J.C. Allwright. On min-max model-based predi tive ontrol. In D. Clarke,

editor, Advan es in Model-Based Predi tive Control, pages 415426. Oxford
University Press, 1994.

[9 A. Amato, M. Mattei, and S. S ala. HIRM design hallenge presentation do -

ument: the LQ methods and eigenstru ture assignment approa h. Te hni al
Publi ation TP-088-26, Group for Aeronauti al Resear h and te hnology in
EURope GARTEUR-FM(AG08), 1997.

[10 B. D. O. Anderson and J. B. Moore. Linear Optimal Control. Prenti e Hall,

1992.

[11 A.N. Andry, E.Y. Shapiro, and J.C Chung. Eigenstru ture assignment for linear
systems. IEEE Trans. Aerosp. Ele tron. Syst., AES-19:711729, 1983.

[12 E.Y. Andry, A.N. Shapiro and J.C. Chung. Eigenstru ture assignment of linear
systems. IEEE Transa tions on Automati Control, AC-29:711720, 1983.

[13 P. Apkarian and R.J. Adams. Advan ed gain-s heduling te hniques for un-
ertain systems. IEEE Transa tions on Control System Te hnology, 1997. To
appear.

636
[14 I. Ashworth. Control of a vstol air raft using pseudo-derivative feedba k. PhD
thesis, Engineering Department, Lan aster University, UK, 1995.

[15 S.J. Asseo. De oupling of a lass of nonlinear systems and its appli ation to an
air raft ontrol problem. J. Air raft, 10:739747, 1973.

[16 R. Babuka, M. Setnes, U. Kaymak, and H.R. van Nauta Lemke. Rule base
simpli ation with similarity measures. In Pro eedings of the Fifth IEEE In-
ternational Conferen e on Fuzzy Systems, volume 3, pages 16421647, New
Orleans, USA, 1996.

[17 R. Babuka and H.B. Verbruggen. An overview of fuzzy modeling for ontrol.
Control Engineering Pra ti e, 4(11):15931606, 1996.

[18 G. J. Balas, J. C. Doyle, K. Glover, A. Pa kard, and R. Smith. -Analysis

and Synthesis TOOLBOX, For Use with MATLAB. The Math Works In ., rst
edition, July 1993.

[19 J. Bals. Hyperstable ontroller parametrization and multiobje tive optimiza-

tion for a tive exible stru tures - a ase study. In Dynami and Control of
Stru tures in Spa e II, Pro . of the 2nd International Conferen e on Dynami s
and Control of Stru tures in Spa e, Craneld, pages 481496. Editors: C.L.
Kirk, P.C. Hughes, 1993.

[20 J. Bals, W. Fi hter, and M. Surauer. Optimization of magneti and angular-

momentum ontrol for low earth orbit satellites. In 3rd ESA International
Conferen e on Spa e raft Guidan e, Navigation and Control Systems, ESTEC,
Noordwijk, NL, 1996.

[21 G. Baumgarten, J.J. Bu hholz, and W. Heine. A new re onguration on ept

for ight ontrol systems. In Pro eedings of the AIAA Guidan e, Navigation
and Control Conferen e, Baltimore (Paper 95-3176), pages 919, 1995.

[22 J.-M. Baus hat. On the appli ation of a nonlinear in-ight simulation te hnique.
In Pro eedings of the First European Control Conferen e (ECC 91), Grenoble,
Volume 3, pages 24152422, 1991.

[23 A.J. Beaumont, A.D. Noble, and A.S. Mer er. Predi tive ontrol of transient
engine testbeds. In Pro eedings, Control 88, Oxford. Institution of Ele tri al
Engineers, London, 1988.

[24 P. Bendotti and M. M'Saad. A skid-to-turn missile autopilot design: the gen-
eralised predi tive adaptive ontrol approa h. Int.Jnl. Adaptive Control and
Signal Pro essing, 7:1331, 1993.

[25 S. Bennani and G. Looye. RCAM design hallenge presentation do ument: the
-synthesis approa h. Te hni al Publi ation TP-088-11, Group for Aeronauti al
Resear h and te hnology in EURope GARTEUR-FM(AG08), 1997.

[26 H.R. Berenji, R.N. Lea, Y. Jani, P.S. Khedar, A. Malkani, and J. Hoblit. Spa e
shuttle attitude ontrol by reinfor ement learning and fuzzy logi . In 2nd IEEE
Conferen e on Fuzzy Systems, pages 13961401, San Fransis o, CA, 1993.

[27 R. Bitmead, M. Gevers, and V. Wertz. Optimal Adaptive Control  The Think-
ing Man's GPC. Prenti e-Hall, rst edition, 1990.

[28 R.L. Blight, J.D. Dailey and D. Gangsaas. Pra ti al ontrol law design for
air raft using multivariable te hniques. Int. Journal of Control, 59:93137,
1994.

637
[29 K.A. Bordignon and W.C. Durham. Closed-form solutions to onstrained on-
trol allo ation problem. Journal of Guidan e, Control and Dynami s, 18(5),
1995.

[30 P. Borne. Automatique - Analyse et rgulation des pro essus industriels, Tome
I: Rgulation ontinue. Editions Te hnip, Paris, 1993.

[31 G. Bouwer, W.V. Gruenhagen, and H.J. Pausder. Model following ontrol
for tailoring handling qualities - ACT experien e with ATTHeS. In AGARD
CP-560: A tive Control Te hnology: Appli ations and Lessons Learned, Turin,
1995.

[32 S. Boyd. Robust ontrol tools: Graphi al user-interfa es and LMI algorithms.
Systems, Control and Information, 38(3):111117, Mar h 1994. Spe ial issue
on Numeri al Approa hes in Control Theory.

[33 A. Bradshaw and J.M. Counsell. Design of autopilots for high performan e
missiles. Pro . Instn. Me h. Engrs., Part I, 206(I2):7584, 1992.

[34 R.W. Bro kett. Feedba k invariants for nonlinear systems. Pro . VII IFAC
Congress, Helsinki, pages 11151120, 1978.

[35 R. Bro khaus. Flugregelung. Springer Verlag, Berlin, 1994.

[36 M. Brown and C. Harris. Neurofuzzy adaptive modelling and ontrol. Prenti e
Hall, 1994.

[37 R. W. Brumbaugh. Air raft model for the aiaa ontrols design hallenge. AIAA
Journal of Guidan e, Control and Dynami s, 17, 1994.

[38 J.J. Bu hholz. LADICO - lateral/dire tional ontrol of an air raft. Te hni al
Report IB 111-91/32, DLR, 1991.

[39 J.J. Bu hholz. Time delay indu ed by ontrol surfa e rate saturation. Zeits hrift
fuer Flugwissens haft und Weltraumfors hung (ZfW), Bd. 17, 1993.

[40 J.J. Bu hholz, J.M. Baus hat, K.-U. Hahn, and H.J. Pausder. ATTAS &
ATTHeS in-ight simulators. In AGARD CP-577: Flight Simulation - Where
are the Challenges? Brauns hweig, 1996.

[41 D.J. Bugajski, D.F Enns, and M.R. Elgersma. A dynami inversion based
ontrol law with appli ation to the HARV. AIAA Guidan e and Control Conf.,
Portland, OR., AIAA-90-3407, 1990.

[42 C.I. Byrnes and A. Isidori. A survey on re ent developments in nonlinear ontrol
theory. Pro . IFAC Symp., Bar elone, pages 287291, 1985.

[43 G. Cal ev, R. Gorez, and I. Dumitra he. A Popov type approa h to stability
analysis of fuzzy ontrol systems. In Pro eedings of Fourth European Congress
on Intelligent Te hniques and Soft Computing, volume 1, pages 37, Aa hen,
Germany, September 1996.

[44 C. Champetier, J.F. Magni, and P. Apkarian. Algebrai Loop Transfer Re-
overy : an appli ation to the design of a heli opter output feedba k ontrol
law. In Pro . AIAA Conf. on Guidan e Navigation and Control, Boston, Mas-
sa husetts, II:12771282, August 1989.

[45 R.Y. Chiang and M.G. Safanov. Robust Control Toolbox User's Guide. The
Mathworks In ., 24 Prime Park Way, Nati k, Mass. 01760, USA., August 1992.

[46 M. Chilali and P. Gahinet. H1 design with pole pla ement onstraint : A LMI
approa h. I.E.E.E. Trans. on Automat. Control, 341(3), 1996.

638
[47 D.W. Clarke. Advan es in model-based predi tive ontrol. In D. Clarke, editor,
Advan es in Model-Based Predi tive Control, pages 321. Oxford University
Press, 1994.

[48 J.M. Counsell. Optimum and safe ontrol algorithm OSCA for modern missile
autopilot design. PhD thesis, Engineering Department, Lan aster University,
UK, 1992.

[49 J. Daafouz, D. Arzelier, G. Gar ia, and J. Bernussou. RCAM design hallenge
presentation do ument: the lyapunov approa h. Te hni al Publi ation TP-088-
14, Group for Aeronauti al Resear h and te hnology in EURope GARTEUR-
FM(AG08), 1997.

[50 R.L. Dailey. Le ture notes for the workshop on H1 and  methods for robust
ontrol. IEEE Conferen e on De ision and Control, AC-26:7593, 1991.

[51 B. Dang Vu and O.L. Mer ier. A nonlinear ight ontrol law for air-to-ground
gunnery. AGARD Conf. on Integration of Fire Control, Flight Control and
Propulsion Control Systems, AGARD-CP-349:21.121.10, 1983.

[52 J.M. de la Cruz, P. Ruiprez, and J. Aranda. RCAM design hallenge presen-
tation do ument: an eigenstru ture assignment approa h. Te hni al Publi a-
tion TP-088-22, Group for Aeronauti al Resear h and te hnology in EURope
GARTEUR-FM(AG08), 1997.

[53 I. Delgado and A. Martinez et al. Robust ontrol te hniques tutorial do u-

ment. Te hni al Publi ation TP-088-7, Group for Aeronauti al Resear h and
te hnology in EURope GARTEUR-FM(AG08), 1997.

[54 K.H. Doets h. The time ve tor method for stability investigations. ARC R&M,
No. 2954, 1953.

[55 C. Dll, J.F. Magni, and Y. Le Gorreg. RCAM design hallenge presenta-
tion do ument: a modal multi-model ontrol approa h. Te hni al Publi a-
tion TP-088-12, Group for Aeronauti al Resear h and te hnology in EURope
GARTEUR-FM(AG08), 1997.

[56 P. Dorato. Robust ontrol. In IEEE Press, 1987.

[57 J. C. Doyle. Guaranteed margins for LQG regulators. IEEE Transa tions on
Automati Control, 23, 1978.

[58 J.C. Doyle. Analysis of feedba k systems with stru tured un ertainties. IEE
Pro eedings, Part D, Control Theory and Appli ations, 129:242250, 1982.

[59 J.C. Doyle. Synthesis of robust ontrollers and lters. In Pro eedings of the
22th IEEE Conferen e on De ision and Control, pages 109114, San Antonio,
TX, 1983.

[60 J.C. Doyle. Stru tured un ertainty in ontrol system design. In Pro eedings
of the 24th IEEE Conferen e on De ision and Control, pages 260265, Fort
Lauderdale, FL, 1985.

[61 J.C. Doyle, B.A. Fran is, and A.R. Tannenbaum. Feedba k Control Theory.
Maxwell Ma millan Interantional Edition, 1992.

[62 J.C. Doyle, K. Glover, and P. Khargonekar. State-spa e solutions to standard

H2 and H1 ontrol problems. Pro eedings of the Ameri an Control Conferen e,
pages 16911696, 1988.

639
[63 J.C. Doyle, K. Glover, P.P. Khargonekar, and B.A. Fran is. State-spa e solu-
tions to standard H2 and H1 ontrol problems. IEEE Transa tions on Auto-
mati Control, 17(AC-34):831847, August 1989.

[64 J.C. Doyle and G. Stein. Multivariable feedba k design: Con epts for a las-
si al/modern synthesis. IEEE Transa tions on Automati Control, 31(1):416,
1981.

[65 J.C. Doyle, J. E. Wall, and G. Stein. Performan e and robustness analysis for
stru tured un ertainty. Pro eedings IEEE Conferen e on De ision and Control,
pages 629636, 1982.

[66 D. Driankov, H. Hellendoorn, and M. Reinfrank. An introdu tion to fuzzy

ontrol. Springer Verlag, 1993.

[67 H. Duda, G. Bouwer, J.M. Baus hat, and K.U. Hahn. RCAM design hallenge
presentation do ument: a model following ontrol approa h. Te hni al Publi-
ation TP-088-25, Group for Aeronauti al Resear h and te hnology in EURope
GARTEUR-FM(AG08), 1997.

[68 D. Dumur and P. Bou her. Predi tive ontrol appli ation in the ma hine-tool
eld. In D. Clarke, editor, Advan es in Model-Based Predi tive Control, pages
471482. Oxford University Press, 1994.

[69 W.C. Durham and K.A. Bordignon. Multiple ontrol ee tor rate limiting.
Journal of Guidan e, Control and Dynami s, 19(1), 1996.

[70 General Dynami s. AFTI/F-16 development and integration program, DFCS

phase nal te hni al report, AFWAL-TR-843008. General Dynami s, Ft.
Worth, 1984.

[71 D. Enns, D. Bugajski, R. Hendri k, and G. Stein. Dynami inversion: An

evolving methodology for ight ontrol design. AGARD CP 560, 1994.

[72 B. Es ande. HIRM design hallenge presentation do ument: the NDI/LQ ap-
proa h. Te hni al Publi ation TP-088-30, Group for Aeronauti al Resear h and
te hnology in EURope GARTEUR-FM(AG08), 1997.

[73 B. Es ande. HIRM design hallenge presentation do ument: The robust in-
verse dynami s estimation approa h. Te hni al Publi ation TP-088-26, Group
for Aeronauti al Resear h and te hnology in EURope GARTEUR-FM(AG08),
1997.

[74 B. Etkin. Dynami s of ight - stability and ontrol, 2nd edition. John Wiley &
Sons, USA, 1982.

[75 P. L. Falb and W. A. Wolovi h. De oupling in the design and synthesis of

multivariable ontrol systems. IEEE Trans. Auto. Control, AC-12(6):651659,
1967.

[76 L.F. Faleiro and R.W. Pratt. Multi-obje tive eigenstru ture assignment in the
design of ight ontrol systems. In Pro . IFAC word ongress, San Fran is o,
July 1996.

[77 L.F. Faleiro and R.W. Pratt. RCAM design hallenge presentation do u-
ment: The eigenstru ture assignment approa h. Te hni al Publi ation TP-088-
18, Group for Aeronauti al Resear h and te hnology in EURope GARTEUR-
FM(AG08), 1997.

[78 C. Favre. Modern ight ontrol system, a pilot partner towards better safety.
ISASTI'96, Jakarta, Indonesia, June 1996.

640
[79 R. Finsterwalder. A parallel oordinate editor as a visual de ision aid in
a multiobje tive on urrent ontrol engineering environment. In Pro . 5th
IFAC/IMACS Symposium on Computer Aided Design in Control Systems, Uni-
versity of Wales, Swansea, UK, pages 118122. Pergamon Press, Oxford, 1991.

[80 M. Fliess. A new approa h to the nonintera ting ontrol problem in nonlinear
system theory. Pro . 23rd Allerton Conf., pages 123129, 1985.

[81 W. Foag. Regelungste hnis he Konzeption einer aktiven PKWFederung mit

preview . PhD thesis, Ruhr-Universitt Bo hum, VDIVerlag, 1990.

[82 G.F. Franklin, J.D. Powell, and A. Emami-Naeini. Feedba k Control of Dynami
Systems. Addison-Wesley, third edition, 1994.

[83 E. Freund. Design of time-variable multivariable systems by de oupling and by

the inverse. IEEE Trans. Auto. Control, AC-16(2):183185, 1971.

[84 P. Gahinet and A. Nemirovskii. A pa kage for manipulating and solving LMI's.
Tutorial Workshop, C.D.C. San Antonio, 1993.

[85 G. Gar ia and J. Bernussou. Pole assignment for un ertain systems in a spe ied
disk by state feedba k. I.E.E.E. On Trans. on Automati Control, 40(1), 1995.

[86 G. Gar ia, J. Bernussou, and D. Arzelier. Robust stabilization of dis rete time
linear systems with norm bounded time varying un ertainty. Systems and Con-
trol Letters, 22, 1994.

[87 G. Gar ia, J. Bernussou, and D. Arzelier. A LMI solution for disk pole lo ation
with H2 guaranteed ost. In European Control Conferen e, Roma - Italy, 1995.
[88 J. Gautrey. RCAM design hallenge presentation do ument: the lassi al on-
trol approa h. Te hni al Publi ation TP-088-13, Group for Aeronauti al Re-
sear h and te hnology in EURope GARTEUR-FM(AG08), 1997.

[89 A. Tonon. HIRM design hallenge presentation do ument: the lassi al ontrol
approa h. Te hni al Publi ation TP-088-24, Group for Aeronauti al Resear h
and te hnology in EURope GARTEUR-FM(AG08), 1997.

[90 F.W. Gembi ki and Y.Y. Haimes. Approa h to performan e and sensitivity
multiobje tive optimization: The goal attainment method. IEEE Trans. AC
20, pages 769771, 1975.

[91 J.C. Geromel, P.L.D. Peres, and S.R. Souza. H2 guaranteed ost for un ertain
ontinuous-time linear systems. Systems and Control Letters, 19, 1992.

[92 K. Glover and J.C. Doyle. State-Spa e Formulae for all Stabilizing Controllers
that satisfy an H1 Norm Bound and Relations to Risk Sensitivity. Systems &
Control Letters, 11:167172, 1988.

[93 P.J. Gold and J.B. Dryfoos. Design and pilot evaluation of the
RAH-66 Coman he sele table ontrol modes. In Pro eedings of Ameri an Heli-
opter So iety Conferen e, San Fran is o, 1993.

[94 R.M.P. Goverde. Robust Flight Control System Design Using H1 =-Synthesis.
NLR Te hni al Report CR 95356 L, National Aerospa e Laboratory NLR, Am-
sterdam, 1995.

[95 A Gra e. MATLAB Optimisation Toolbox User's guide. The MathWorks In .

[96 M. Green and D.J.N. Limebeer. Linear Robust Control. Prenti e Hall, rst
edition, 1995.

641
 [97 G. Grbel. The ANDECS CACE Framework. IEEE Control Systems, 15(2):8
13, April April, 1995.

[98 G. Grbel, J. Bals, R. Finsterwalder, H.-D. Gramli h, G.and Joos, and M. Ot-
ter. Computer-integrated ontrol-dynami s  design experimentation by an-
de s. In ESA Workshop Spa e raft Guidan e Navigation and Control Systems
Software for Design and Implementation, Noordwijk, 29.9.- 1.10., 1992.

[99 G. Grbel, R. Finsterwalder, G. Gramli h, H.D. Joos, and S. Lewald. ANDECS:

A Computation Environment for Control Appli ations of Optimization, volume
115 of International Series of Numeri al Mathemati s (ISNM), pages 237254.
R. Bulirs h, D. Kraft, eds. Birkhuser, Basel, Boston, Berlin, 1994.

[100 G. Grbel and H.D. Joos. Performan e-Robust Design via a Multi rite-
ria/Multimodel Approa h - a Flight Control Case Study, volume 34, pages 173
195. NATO ASI Series F, Springer Verlag, r.f. urtain edition, 1986.

[101 G. Grbel and H.D. Joos. RASP and RSYST - two omplementary program
libraries for on urrent ontrol engineering. In Pro . 5th IFAC/IMACS Sym-
posium on Computer Aided Design in Control Systems, University of Wales,
Swansea, UK, pages 101106. Pergamon Press, Oxford, 1991.

[102 G. Grbel and G. Kreisselmeier. Systemati omputer aided ontrol design. In

AGARD Conf. Pro . No. 321, AGARD Guidan e and Control Panel Symp. on
Advan es in Guidan e and Control Systems, Lissabon, 1983. AGARD Le ture
Series No. 12.8 Computer Aided Design and Analysis of Digital Guidan e and
Control Systems, Stuttgart/Athen/Paris.

[103 G. Grbel et al. Design hallenge evaluation report. Te hni al Publi a-

tion TP-088-34, Group for Aeronauti al Resear h and te hnology in EURope
GARTEUR-FM(AG08), 1997.

[104 W.V. Gruenhagen. A high bandwidth ontrol system for a heli opter in-ight
simulator. International Journal of Control, 59(1):239261, 1994.

[105 P. Guer het and J.L. Estival. Line of sight guidan e law by predi tive fun -
tional ontrol for high-velo ity short-range ta ti al missile. In Pro eedings First
European Control Conferen e, Grenoble, 1991.

[106 F.Y. Hadaegh, E. Mettler, and C.F. Lin. Autonomous spa e raft guidan e and
ontrol. In AIAA Guidan e Navigation and Control Conferen e, AIAA 96-3924,
San Diego, CA, 1996.

[107 P.G. Hamel. Te hniques for model identi ation and design of model following
ontrol systems. In Pro eedings of International Symposium on Stability, pages
257283, 1987.

[108 P.G. Hamel. Aerospa e vehi le modelling requirements for high bandwidth
ight ontrol. In Aerospa e Vehi le Dynami s and Control, pages 132. Oxford
University Press, 1994.

[109 P.G. Hamel and R.V. Jategaonkar. The evolution of ight vehi le system iden-
ti ation. Journal of Air raft, 33(1):928, 1996.

[110 R. Hanus, M. Kinnaert, and J.L. Henrotte. Conditioning te hnique, a general

anti-windup and bumpless transfer method. Automati a, 23(6):729739, 1987.

[111 L.H. Harrison, P.J. Saunders, and P.J. Sara eni. Arti ial intelligen e and
expert systems for avioni s. In AIAA-IEEE 12th Digital Avioni s Systems
Conferen e, pages 167172, Fort Worth, TX, 1993.

642
[112 A. Helmersson. Methods for robust gain-s heduling. Te hni al report, PhD
Thesis, Linkping University, Sweden, 1995.

[113 F. Hens hel and S. Chetty. Flight ontrol system design for an in-ight simu-
lator. Journal of Guidan e, Control, and Dynami s, 12(3):351356, 1989.

[114 L.P. Holmblad and J.J. Ostergaard. Control of a ement kiln by fuzzy logi .
In M.M. Gupta and E. San hez, editors, Fuzzy Information and De ision Pro-
esses, pages 389399. North Holland, 1982.

[115 D.J.F. Hopper. A tive ontrol of v/stol air raft. PhD thesis, Engineering De-
partment, Salford University, UK, 1990.

[116 C.H. Houpis, R.R. Sating, S. Rasmussen, and S. Skeldon. Quantitative feedba k
theory te hnique and appli ations. Int. J. on Control, Spe ial Issue on Air raft
Flight Control, 59:3970, 1994.

[117 D. Hoyle, R. Hyde, and D.J.N. Limebeer. An H1 Approa h to Two-Degree-Of-

Freedom Design. Pro eedings of the IEEE CDC, pages 15811585, De ember
1991.

[118 M. Huzmezan and J.M. Ma iejowski. A development spa e for model based pre-
di tive ontrol. In 7th Symposium on Computer Aided Control System Design,
Ghent,Belgium. IFAC, 1997. Submitted.

[119 M. Huzmezan and J.M. Ma iejowski. RCAM design hallenge presentation

do ument: the model based predi tive ontrol approa h. Te hni al Publi a-
tion TP-088-20, Group for Aeronauti al Resear h and te hnology in EURope
GARTEUR-FM(AG08), 1997.

[120 R.A. Hyde. H1 Aerospa e Control Design. A VSTOL Flight Appli ation. Ad-
van es in Industrial Control Series. Springer-Verlag, 1995.

[121 The MathWorks In . Matlab user's guide. July 1993.

[122 A. Inselberg. Parallel oordinates for visualizing multi-dimensional geometry.

In Pro eedings of Computer Graphi s International '87, 1987.

[123 J. Irving. Flight ontrol law pro ess model for use in the garteur a tion group.
Te hni al Publi ation TP-088-2, Group for Aeronauti al Resear h and te hnol-
ogy in EURope GARTEUR-FM(AG08), 1995.

[124 A. Isidori. Nonlinear ontrol systems, in: Communi ations and Control Engi-
neering, 2nd Edition. Springer-Verlag, Berlin, 1989.

[125 A. Isidori, A.J. Krener, Giorgi Gori, C., and S. Mona o. Nonlinear de oupling
via feedba k : a dierential geometri approa h. IEEE Trans. Auto. Control,
AC-26:331345, 1981.

[126 H.G. Ja ob. Re hnergestuetzte Optimierung statis her und dynamis her Sys-
teme. Springer-Verlag, Berlin Heidelberg New York, 1982.

[127 J.S.R. Jang and C.T. Sun. Neuro-fuzzy modeling and ontrol. Pro eedings of
IEEE, 83(4):378406, 1995.

[128 H.D. Joos. Informationste hnis he Behandlung des mehrzieligen optimierungs-

gesttzten regelungste hnis hen Entwurfs. PhD thesis, Universitt Stuttgart,
Prof. R. Rhle, 1993. Forts hrittberi hte VDI, Reihe 20: Re hneruntersttzte
Verfahren, Nr. 90.

643
[129 H.D. Joos. ANDECS-multi riteria optimization of a Fuzzy-Logi ontroller for
an aerodynami ally unstable air raft in omparison to a linear robust ontrol
design. In IFAC Conferen e on Integrated Systems Engineering, Baden-Baden,
pages 275280, 1994.

[130 H.D. Joos. RCAM design hallenge presentation do ument: Multi-obje tive pa-
rameter synthesis (MOPS). Te hni al Publi ation TP-088-16, Group for Aero-
nauti al Resear h and te hnology in EURope GARTEUR-FM(AG08), 1997.

[131 R. E. Kalman. When is a linear ontrol system optimal? Trans. ASME Ser.
D: J. Basi Eng., 86, 1964.

[132 I. Kaminer, R.A. Benson, E. Coleman, and Y. Ebrahimi. Design of integrated

pit h axis for autopilot/autothrottle and integrated lateral axis autopilot/yaw
damper for NASA TSRV airplane using integral LQG methodology. NASA-
CR 4268, National Aeronauti s and Spa e Administration, Langley Resear h
Center, 1990.

[133 J. Kaustky, N.K. Ni hols, and P. Van Dooren. Robust pole assignment in linear
state feedba k. Int. Journal of Control, 41(5):11291155, 1985.

[134 P.P. Khargonekar, I.R. Petersen, and K. Zhou. Robust stabilization of un er-
tain linear systems: Quadrati stabilizability and H1 ontrol theory. I.E.E.E.
Trans. on Automat. Control, 35(pp 356-361), 1990.

[135 H. Kimura. Pole assignment by gain output feedba k. IEEE Transa tions on
Automati Control, AC-20:509516, 1975.

[136 G.J. Klir and B. Yuan. Fuzzy sets and fuzzy logi . Prenti e Hall, 1995.

[137 B. Kouvaritakis, J.A. Rossiter, and A.O.T. Chang. Stable generalized predi tive
ontrol: An algorithm with guaranteed stability. Pro eedingsIEE  Part D,
139(4):349362, 1992.

[138 G. Kreisselmeier and R. Steinhauser. Systematis he Auslegung von Re-

glern dur h Optimierung eines vektoriellen Gtekriteriums. Regelungste hnik,
27(3):7679, 1979.

[139 A.J. Krener. On the equivalen e of ontrol systems and linearization of nonlin-
ear systems. SIAM J. Control and Opt., 11:670676, 1973.

[140 H. Kwakernaak and R. Sivan. Linear Optimal Control Systems. Wiley-

Inters ien e, 1972.

[141 H. Kwakernaak and R. Sivan. The maximally a hievable a ura y of linear

optimal regulators and linear optimal lters. IEEE Transa tions on Automati
Control, 17, 1972.

[142 W.A. Kwong, K.M. Passino, and E.G. Laukonen. Expert supervision of fuzzy
learning systems for fault tolerant air raft ontrol. Pro eedings of IEEE, 83(3),
1995.

[143 P.F. Lambre hts. Addendum to the RCAM design hallenge material. Memo-
rendum M-088-6, Group for Aeronauti al Resear h and te hnology in EURope
GARTEUR-FM(AG08), 1996.

[144 P.F. Lambre hts, J.C. Terlouw, S. Bennani, and M. Steinbu h. Parametri
Un ertainty Modeling using LFTs. In Pro eedings of the Ameri an Control
Conferen e, pages 267272, San Fran is o, California, 1993.

644
[145 P.F. Lambre hts et al. Robust ight ontrol design hallenge problem formula-
tion and manual: the resear h ivil air raft model (RCAM). Te hni al Publi-
ation TP-088-3, Group for Aeronauti al Resear h and te hnology in EURope
GARTEUR-FM(AG08), 1997.

[146 A.A. Lambregts. Operational aspe ts of the integrated verti al ight path and
speed ontrol system. AIAA Paper 83-1420, 1983.

[147 A.A. Lambregts. Verti al ight path and speed ontrol autopilot design using
total energy prin iples. Te hni al report, AIAA 83-2239, 1983.

[148 A.A. Lambregts. Sheets of olloquia at DUT on automati ight ontrol. Te h-

ni al report, Delft University of Te hnology, 1996.

[149 S. H. Lane and R.F. Stengel. Flight ontrol design using nonlinear inverse
dynami s. Automati a, 24(4):471483, 1988.

[150 Y. Le Gorre , J.F. Magni, C. Dll, and C. Chiappa. A modal multimodel

ontrol design approa h applied to air raft autopilot design. To appear in the
AIAA Journal of Guidan e Control and Dynami s, 1997.

[151 J.H. Lee and Z.H. Yu. Tuning of model predi tive ontrollers for robust perfor-
man e. Computers in Chemi al Engineering, 18(1):1537, 1994.

[152 N.A. Lehtomaki, N.S. Sandell, and M. Athans. Robustness result in linear
quadrati gaussian based multivariable ontrol designs. IEEE Transa tions on
Automati Control, AC-26:7592, 1981.

[153 A. Lewald. Ein neuartiges Verfahren zur numeris hen Bere hnung zeitoptimaler
Robotersteuerungen. PhD thesis, Ruhr-Universitt Bo hum, 1994. Forts hrit-
tberi hte VDI, Reihe 8: Me"s-, Steuerungs- und Regelungste hnik, Nr. 396.

[154 D.J.N. Limebeer, E.M. Kasenally, and J.D. Perkins. On the design of robust
two degree of freedom ontrollers. Automati a, 29(1):157168, 1993.

[155 C.F. Lin. Advan ed Control Systems Design. Prenti e-Hall, N.J., 1994.

[156 D.J. Linse and R.F. Stengel. Identi ation of aerodynami oe ients using
omputational neural networks. In 30th AIAA Aerospa e S ien es Meeting &
Exhibit, AIAA 92-0172, Reno, NV, 1992.

[157 G. Looye and S. Bennani. Des ription and analysis of the resear h ivil air raft
model (r am). Te hni al report, DUT, Aerospa e Engineering, se t. Stability
and Control, Delft, June 1996.

[158 G. Looye, G. Moormann, A. Varga, and S. Bennani. Post-Design Stability

Robustness Assessment of the RCAM Controller Design Entries. Te hni al
Publi ation TP-088-35, Group for Aeronauti al Resear h and te hnology in
EURope GARTEUR-FM(AG08), 1997.

[159 J.M. Ma iejowski. Multivariable Feedba k Design. Ele troni Systems Engi-
neering Series. Addisson-Wesley Publishing Company, Wokingham, England,
1989.

[160 J.F. Magni and A. Manouan. Robust ight ontrol design by eigenstru ture
assignment. in Pro . of IFAC Symposium on Robust Control, Rio de Janeiro,
Brasil, September 1994.

[161 Y. Magni, J.F.and Le Gorre . An observer based multimodel ontrol design

approa h. Submitted for publi ation to the Asian Control Conferen e, Seoul,
Corea, 1997.

645
[162 E.H. Mamdani. Appli ations of fuzzy algorithms for ontrol of simple dynami
plant. Pro eedings IEE, 121:15851588, 1974.

[163 J. Markerink and S. Bennani. HIRM design hallenge presentation do ument:

the -synthesis approa h. Te hni al Publi ation TP-088-29, Group for Aero-
nauti al Resear h and te hnology in EURope GARTEUR-FM(AG08), 1997.

[164 D.C. M Farlane and K. Glover. Robust Controller Design Using Normalized
Coprime Fa tor Plant Des riptions. Le ture Notes in Control and Information
S ien es. Springer-Verlag, rst edition, 1990.

[165 D. M Leon. Automati Flight Control Systems. Prenti e Hall, New York, 1990.

[166 D. M Ruer, Ashkenas, and Graham. Air raft Dynami s and Automati Control.
Prin eton University Press, 1973.

[167 D. M Ruer and H.R. Jex. A review of quasi-linear pilot models. IEEE Trans-
a tions on Human Fa tors in Ele troni s, 8(3):231249, 1967.

[168 P. K. A Menon, M. E. Badgett, and R.A. Walker. Nonlinear ight test traje tory
ontrollers for air raft. AIAA Guidan e and Control Conf., Snow Mass, CO,
AIAA-85-1890, 1985.

[169 O. Mer ier. Realisation de regulateurs de poursuite ameliores pour le pilotage

des avions a grande manoeuvrabilite. In 35th AGARD GCP Symposium Ad-
van es in Guidan e and Control Systems , 1982.

[170 G. Meyer and L. Ci olani. Appli ation of nonlinear system inverses to automati
ight ontrol designs system on epts and ight evaluations. Theory and Appli-
ation of Optimal Control in Aerospa e Systems, AGARD AG251:10.110.29,
1981.

[171 G. Meyer et al. Nonlinear ontroller design for ight ontrol systems. In Pro .
IFAC Symp., Nonlinear Control Systems Design, Capri, pages 136141, 1989.

[172 D.D. Moerder and A.J. Calise. Convergen e of a numeri al algorithm for al u-
lating optimal output feedba k gains. IEEE Transa tions on Automati Con-
trol, 30(9):900903, 1985.

[173 M. Morari. Model predi tive ontrol: multivariable ontrol te hnique of hoi e
in the 1990's? In D. Clarke, editor, Advan es in Model-Based Predi tive Control,
pages 2237. Oxford University Press, 1994.

[174 E. Mos a. Optimal, Predi tive and Adaptive Control. Prenti e-Hall, rst edition,
1995.

[175 P.R. Motyka, E.G. Rynaski, and P.A. Reynolds. Theory and ight veri ation
of the tifs model-following system. Journal of Air raft, 9(5):347353, 1972.

[176 E.A.M. Muir. Appli ation of robust inverse dynami s estimation to the ontrol
of a thrust ve toring ghter air raft. PhD thesis, Engineering Department,
Lan aster University, UK, 1996.

[177 E.A.M. Muir et al. Robust ight ontrol design hallenge problem formulation
and manual: The high in iden e resear h model (HIRM), version 1. Te hni-
al Publi ation TP-088-4, Group for Aeronauti al Resear h and te hnology in
EURope GARTEUR-FM(AG08), August 1997.

[178 V. Mukhopadhyay. Stability robustness improvement using onstrained op-

timisation te hniques. AIAA Journal of Guidan e, Navigation and Control,
10:172177, Mar h 1987.

646
[179 S.S. Mulgund and R.F. Stengel. Air raft ight ontrol in wind shear using
partial dynami inversion. In Pro eedings of the Ameri an Control Conferen e,
pages 400404, San Fran is o, California, June 1993.

[180 S.S. Mulgund and R.F. Stengel. Target pit h angle for the mi roburst es ape
maneuver. Journal of Air raft, 30(6):826832, Nov.De . 1993.

[181 W. Murray and M.L. Overton. A proje ted lagragian algorithm for nonlinear
minmax optimization. SIAM J.S i.Sta.Comp. 1, pages 345370, 1980.

[182 M.R. Napolitano, C.I. Chen, and S. Naylor. Air raft failure dete tion and
identi ation using neural networks. AIAA Journal of Guidan e, Control, and
Dynami s, 16(6):9991009, 1993.

[183 M.R. Napolitano and M. Kin heloe. On-line learning neural-network ontrollers
for autopilot systems. AIAA Journal of Guidan e, Control, and Dynami s,
33(6):10081015, 1995.

[184 H. Nijmeijer and A. Van Der S haft. Nonlinear dynami al ontrol systems.
Springer-Verlag, New York, 1990.

[185 R.G.E. Nijo, C.W. S herer, and S. Bennani. Appli ation of LPV ontrol with
full blo k s alings for a high performan e ight ontrol system. Sele ted Topi s
in Identi ation, Modelling and Control, 9, De ember 1996.

[186 M. Otter. DSblo k: A neutral des ription of dynami systems Version 3.2.
DLRIB 55192/11, DLR Deuts he Fors hungsanstalt fr Luft- und Raum-
fahrt, D-82234 Oberpfaffenhofen, 1992.

[187 A. Pa kard. A new design method for ontroller order redu tion, a -
perspe tive. Preprint, 1995.

[188 A. Pa kard and J.C. Doyle. The omplex stru tured singular value. Automati a,
29(1):71109, 1993.

[189 A. Pa kard, J.C. Doyle, and G. Balas. Linear, multivariable robust ontrol with
a -perspe tive. Trans.ASME/Journ. Dyn. Syst. Measurement and Control,
115:426438, 1993.

[190 Paddison, J.E. and Goodall, R.M. and Bals, J. and Grbel, G. Multi-obje tive
design study for a Maglev suspension ontroller using the databased ANDECS-
MATLAB environment. In IEEE/IFAC CACSD'94 Symposium Tu son, Mar h
7-9, pages 239246, 1994.

[191 G. Papageorgiou, K. Glover, R.A. Hyde, and G.D. Wood. HIRM design hal-
lenge presentation do ument: the H1 loop-shaping approa h. Te hni al Publi-
ation TP-088-32, Group for Aeronauti al Resear h and te hnology in EURope
GARTEUR-FM(AG08), 1997.

[192 G. Papageorgiou, M. Huzmezan, K. Glover, and J. Ma iejowski. Combined

MBPC/ H1 autopilot for a ivil air raft. Ameri an Control Conferen e,New
Mexi o, USA, 1997. Submitted.

[193 I.R. Petersen. A stabilization algorithm for a lass of un ertain linear systems.
Systems and Control Letters, 8, 1987.

[194 R. Phelan. Automati ontrol systems. Cornell University Press, 1977.

[195 C. Phillips, C.L. Karr, and G. Walker. Heli opter ight ontrol with fuzzy
logi and geneti algorithms. Engineering Appli ations of Arti ial Intelligen e,
9(2):175184, 1996.

647
[196 W.A. Porter. Diagonalisation and inverses for nonlinear systems. Int. J. Con-
trol, 10:252264, 1970.

[197 D.M. Prett and C.E. Gar ia. Fundamental Pro ess Control. Butterworths Series
in Chemi al Engineering, rst edition, 1988.

[198 Y. Ramamurthi, P.B. Sistu, and B.W. Bequette. Control relevant dynami data
re on iliation and parameter estimation. Computers in Chemi al Engineering,
17(1):4159, 1993.

[199 J. Rawlings and K. Muske. The stability of onstrained re eding horizon ontrol.
IEEE Transa tions on Automati Control, 38:15121516, 1993.

[200 J. Reiner, G.J. Balas, and W.L. Garrard. Robust dynami inversion for ontrol
of highly maneuverable air raft. Journal of Guidan e, Control and Dynami s,
18(1), January-February 1995.

[201 P.A. Robinson. The modelling of turbulen e and downbursts for ight simula-
tors. UTIAS report No.339, University of Toronto, 1991.

[202 M. G. Safonov and M. Athans. Gain and phase margin for multiloop LQG
regulators. IEEE Transa tions on Automati Control, 22, 1977.

[203 M.G. Safonov, A.J. Laub, and G.L. Hartmann. Feedba k properties of mul-
tivariable systems: the role and use of the return dieren e matrix. IEEE
Transa tions on Automati Control, 26, 1981.

[204 N. Saravanan, A. Duyar, T.H. Guo, and W.C. Merrill. Modeling spa e shuttle
main engine using feed-forward neural networks. AIAA Journal of Guidan e,
Control, and Dynami s, 17(4):641648, 1994.

[205 Y. Sawaraki and H. Nakayama. The Theory of Multiobje tive Optimization.

A ademi Pres, In ., Mathemati s in S ien e and Engineering, Vol. 176, 1985.

[206 S. S ala et al. Robust ight ontrol literature survey database. Te hni al Pub-
li ation TP-088-6, Group for Aeronauti al Resear h and te hnology in EURope
GARTEUR-FM(AG08), February 1997.

[207 G. S hram et al. RCAM design hallenge presentation do ument: A fuzzy

logi ontrol appro h. Te hni al Publi ation TP-088-15, Group for Aeronauti al
Resear h and te hnology in EURope GARTEUR-FM(AG08), 1997.

[208 J. S huring and R.M.P. Goverde. RCAM design hallenge presentation do -

ument: the -synthesis approa h. Te hni al Publi ation TP-088-19, Group
for Aeronauti al Resear h and te hnology in EURope GARTEUR-FM(AG08),
1997.

[209 W. S hwartz and H.D. Joos. ANDECS omputational experimenting with the
IAVSD ar-suspension ben hmark iltis. In IFAC Conferen e on Integrated Sys-
tems Engineering, Baden-Baden, pages 263267, 1994.

[210 J. A. Sefton and K. Glover. Pole/zero an ellations in the general H1 problem

with referen e to a two blo k design. Systems and Control Letters, 14:295306,
1990.

[211 E.Y. Shapiro, A.N. Andry, and J.C. Chung. Eigenstru ture assignment for
linear systems. IEEE Transa tions on Automati Control, AC-29:12241225,
1984.

[212 S.N. Singh and W.J. Rugh. De oupling in a lass of nonlinear systems by state
variable feedba k. Trans. ASME J. Dyn. Syst. Measur. Contr., 94:323329,
1972.

648
[213 S.N. Singh and A. S hy. Output feedba k nonlinear de oupled ontrol synthesis
and observer design for maneuvering air raft. Int. J. Control, 31:781806, 1980.

[214 S.N. Singh, M. Steinberg, and R.D. DiGirolamo. Nonlinear predi tive ontrol of
feedba k linearizable systems and ight ontrol system design. AIAA Journal
of Guidan e, Control, and Dynami s, 18(5):10231028, 1995.

[215 S. Skogestad and I. Postlethwaite. Multivariable Feedba k Control. John Wiley

& Sons, rst edition, 1996.

[216 J.J.E. Slotine and J.W. Li. Applied nonlinear ontrol. Prenti e Hall, Englewood
Clis, N.J., 1991.

[217 A. Smerlas, I. Postlehwaite, and D. Walker. HIRM design hallenge presentation

do ument: the H1 loop-shaping approa h. Te hni al Publi ation TP-088-
33, Group for Aeronauti al Resear h and te hnology in EURope GARTEUR-
FM(AG08), 1997.

[218 P.R. Smith. Fun tional ontrol law design using exa t non-linear dynami in-
version. AIAA guidan e, navigation and ontrol onferen e pro eedings, AIAA-
94-3516-CP:481489, 1994.

[219 K.M. Sobel and F.J. Lallman. Eigenstru ture assignment for the ontrol of
highly augmented air raft. AIAA Journal of Guidan e, Control and Dynami s,
12:318324, 1989.

[220 K.M. Sobel and E.Y. Shapiro. Appli ation of eigenstru ture assignment to
ight ontrol design: some extensions. AIAA Journal of Guidan e, Control and
Dynami s, 10:7381, 1987.

[221 K.M. Sobel, W. Yu, and F.J. Lallman. Eigenstru ture assignment with gain
suppression using eigenvalue and eigenve tor derivatives. AIAA Journal of
Guidan e, Control and Dynami s, 13, 1990.

[222 A.R.M. Soeterboek. Predi tive Control - A unied approa h. Prenti e Hall,
New York, rst edition, 1992.

[223 A.R.M. Soeterboek, A.F. Pels, H.B. Verbruggen, and G.C.A. van Langen. A
predi tive ontroller for the ma h number in a transoni wind tunnel. IEEE
Control System Magazine, 11(1):6372, 1991.

[224 J. Spee et al. Ca ee overview do ument. Te hni al Publi ation TP-088-

10, Group for Aeronauti al Resear h and te hnology in EURope GARTEUR-
FM(AG08), 1997.

[225 K. Sthl Gunnarson. HIRM design hallenge presentation do ument: the -

synthesis approa h. Te hni al Publi ation TP-088-31, Group for Aeronauti al
Resear h and te hnology in EURope GARTEUR-FM(AG08), 1997.

[226 G. Stein and M. Athans. The LQG/LTR pro edure for multivariable feedba k
ontrol design. IEEE Transa tions on Automati Control, 32(2):105114, 1987.

[227 G. Stein and J.C. Doyle. Beyond singular values and loop shapes. AIAA Journal
of Guidan e Control and Dynami s, 14(1):516, 1991.

[228 M. Steinberg. Potential role of neural networks and fuzzy logi in ight ontrol
design and development. In AIAA Aerospa e Design Conferen e, AIAA-92-
0999, Irvine, CA, 1992.

[229 M. Steinberg. Development and simulation of an F/A-18 fuzzy logi automati

arrier landing system. In 2nd IEEE onferen e on Fuzzy Systems, pages 797
802, San Fransis o, CA, 1993.

649
[230 R.F. Stengel. Towards intelligent ight ontrol. IEEE transa tions on Systems,
Man, and Cyberneti s, 23(6):16991717, 1993.

[231 B.L. Stevens and F.L. Lewis. Air raft Control and Simulation. Wiley-
Inters ien e, 1992.

[232 D.A. Stratton and R.F. Stengel. Real-time de ision aiding: Air raft guidan e for
wind shear avoidan e. IEEE transa tions on Aerospa e and Ele troni Systems,
31(1):117124, 1995.

[233 M. Sugeno, M.F. Grin, and A. Bastian. A fuzzy hierar hi al ontrol of an

unmanned heli opter. In IEEE 3rd Conferen e of Fuzzy Systems, Korea, 1993.

[234 T. Takagi and M. Sugeno. Fuzzy identi ation of systems and its appli ation to
modeling and ontrol. IEEE transa tions on Systems, man, and Cyberneti s,
15(1):116132, 1985.

[235 K. Tanaka and M. Sugeno. Stability analysis and design of fuzzy ontrol sys-
tems. Fuzzy Sets and Systems, 45(2):135156, 1992.

[236 T. Terano, K. Asai, and M. Sugeno. Fuzzy Systems Theory and Its Appli ations.
A ademi Press, Boston, 1992.

[237 J. Terlouw, A. Helmersson, K. Sthl-Gunnarsson, and P.F. Lambre hts. Com-

muni ation handbook, version 2. Te hni al Publi ation TP-088-5, Group
for Aeronauti al Resear h and te hnology in EURope GARTEUR-FM(AG08),
1997.

[238 J.C. Terlouw and P.F. Lambre hts. A Matlab Toolbox for Parametri Un er-
tainty Modelling. NLR Te hni al Publi ation CR 93455 L, National Aerospa e
Laboratory NLR, Amsterdam, 1993.

[239 J.C. Terlouw and P.F. Lambre hts. A MATLAB Toolbox for Parametri Un-
ertainty Modelling. Te hni al Report CR 93455 L, NLR, 1993.

[240 The MathWorks. SIMULINK A program for Simulating Dynami Systems. The
Mathworks Series, 1995.

[241 M.-C. Tsai, E.J.M. Geddes, and I. Postlethwaite. Pole-Zero Can ellations and
Closed-Loop Properties of an H1 Mixed Sensitivity Design Problem. Automat-
i a, 28:519530, 1992.

[242 M. Tu ker and D. Walker. RCAM design hallenge presentation do ument: the
H1 loop-shaping approa h. Te hni al Publi ation TP-088-21, Group for Aero-
nauti al Resear h and te hnology in EURope GARTEUR-FM(AG08), 1997.

[243 J.S. Tyler. The hara teristi s of model-following systems as synthesized by

optimal ontrol. IEEE Transa tions on Automati Control, AC-9(4):485498,
1964.

[244 V.I. Utkin. Variable stru ture systems with sliding modes. IEEE Transa tions
on Automati Control, AC22(2):212222, 1977.

[245 A. Varga, G. Looye, D. Moormann, and G. Grbel. Automated Generation of

LFT-Based Parametri Un ertainty Des riptions from Symboli Air raft Mod-
els. Te hni al Publi ation TP-088-36, Group for Aeronauti al Resear h and
te hnology in EURope GARTEUR-FM(AG08), 1997.

[246 M. Vidyasagar. Control System Synthesis: A Coprime Fa torization Approa h.

MIT Press, 1985.

650
[247 G. Vinni ombe. Measuring Robustness of Feedba k Systems. PhD thesis, De-
partment of Engineering, University of Cambridge, 1993.

[248 C. von Altro k. Towards fuzzy logi standardization. In Pro eedings of Fifth
IEEE International Conferen e on Fuzzy Systems, volume 3, pages 20912093,
New Orleans, USA, September 1996.

[249 P. Voulgaris and L. Valavani. High performan e linear quadrati and H1 de-
signs for a supermaneouverable air raft. AIAA Journal of Guidan e, Control,
and Dynami s, 14(1):157165, 1991.
[250 D.J. Walker. On the Stru tute of a 2-Degrees-Of-Freedom Controller. Interna-
tional Journal of Control, vol. 63, No 6, pages 11051127, 1996.
[251 L.X. Wang. Adaptive Fuzzy Systems and Control. Prenti e-Hall, Englewood
Clis, 1994.

[252 D.G. Ward and R.L. Barron. A self-designing re eding horizon optimal ight
ontroller. In Pro . Ameri an Control Conferen e, 1995.

[253 Y.J.P. Wei. Intelligent Control Law Tuning for AIAA Controls Design Chal-
lenge. Journal of Guidan e, Control, and Dynami s, 17:753758, 1994.

[254 H. Wents her. Design and Analysis of Semi-A tive Landing Gears for Transport
Air raft. PhD thesis, TU Mn hen, 1996. DLRFB 96-11, ISSN 0939-2963.
[255 J.F. Whidborne, I. Postlethwaite, and D.W. Gu. Robust Controller Design
Using H1 Loop-Shaping and the Method of Inequalities. IEEE Trans. Control
Systems Te hnology, 2(4):455461, De ., 1994.
[256 J. H. Wilkinson. The Algebrai Eigenvalue Problem. Monographs on Numeri al
Analysis. Clarendon Press, Oxford, 1965.

[257 D.M.C Willemsen, S. Bennani, and C.W. S herer. Robust LPV ontrol for a
ight ontrol system. Sele ted Topi s in Identi ation, Modelling and Control,
9, De ember 1996.

[258 K. Wise et al. Linear and nonlinear air raft ight ontrol for the AIAA ontrol
design hallenge. Pro eedings of the AIAA Conferen e on Guidan e, Navigation
and Control, 1992.
[259 G.D. Wood. Control of Parameter-Dependent Me hani al Systems. PhD thesis,
Cambridge University, 1996.

[260 F. Wu. Control of linear parameter varying systems. Te hni al report, PhD
thesis, University of California at Berkeley, Berkeley, California, June 1995.

[261 C. Yi and B. de Moor. Robustness analysis and ontrol system design for the
servo system of a ontinuously variable transmission. IEEE Trans. Control
Systems Te hnology, 2(3):183197, September 1994.
[262 P.W. Ying-Jyi. Intelligent ontrol law tuning for AIAA ontrol design hallenge.
AIAA Journal of Guidan e, Control and Dynami s, 17, 1994.
[263 H. Youssef, C.Y. Chiang, and G.R. Yu. On-line LQG-fuzzy approa h to failure
dete tion, isolation, and re onguration of ontrol surfa es. In AIAA Guidan e
Navigation and Control Conferen e, AIAA 96-3799, 1996.
[264 L.A. Zadeh. Fuzzy sets. Information and Control, 8:338353, 1965.

[265 E. Zariou. Robust Model Predi tive Control of pro esses with Hard Con-
straints. Computers Chemi al Engineering, 14(4/5):359371, 1990.

[266 K. Zhou, J.C. Doyle, and K. Glover. Robust and Optimal Control. Prenti e
Hall In ., 1996.

651